date: Wed, 29 Sep 1999 12:09:02 +0200 (MET DST) from: GIORGI FILIPPO subject: chapter 10, draft 1 to: tar 10 site Dear All in the attachment, please find the entire chapter as it now stands. You should consider this as the going official draft. Anything else you might have in your computers (ESPECIALLY IN WORD) does not count. This is now the going plan: 1) You should all read the whole thing (if possible). Especially read your individual sections, but also pay special attention to the "common" sections, i.e. 10.1, 10.2, 10.8. In particular make sure that all that is said there is in tune with what is said in the individual sections. 2) Do not worry about little editorial issues. I and Bruce ar going to go over the text carefully to fix those without changing the meaning of things. At this point there is no time for major "structural" changes. If you have some significant changes you want to make, indicate the original paragraph and the modifications to it. At this point there should not be many of them (but you never know). Direct the changes to all, but I and Bruce will be in charge of putting them in. 3) Let everybody know if there are major issues you want to raise. 4) If you know of work that is not included here but will likely be included in the next draft, make a list and we will include it within the text at the end of the appropriate section (please indicate that too). 5) Hans, please ask your secretary to go over the refs and let us know what is missing, needed etc. 6) Section coordinators, collect all figure captions for your section and please send them to me so I can put them together. 7) Section coordinators, please complete all missing tables and other relevant material and send to Bruce in some ps format. 8) Section coordinators please check with Bruce about figures that he has or that are missing. 9) Give any other comment you might feel is pertinent. All the points 1) through 9) should be completed by end of day thursday (Trieste time). Then on friday I and Bruce will coordinate to produce the final offical draft and put it on the website for the TSU. Couple more days and we should be done. Thansk a lot for your effort in the last weeks. Cheers, Filippo ################################################################ # Filippo Giorgi, Head, # # Physics of Weather and Climate Group # # The Abdus Salam International Centre for Theoretical Physics # # P.O. BOX 586, (Strada Costiera 11 for courier mail) # # 34100 Trieste, ITALY # # Phone: + 39 (040) 2240 425 # # Fax: + 39 (040) 224 163 # # email: giorgi@ictp.trieste.it # ################################################################ 10.1 Introduction This Chapter is a new addition compared to previous IPCC assessment reports. It stems from the increasing need to evaluate regional climate change information for use in impact studies and policy planning. To date, regional climate change information has been characterized by a relatively high level of uncertainty. This is due to the complexity of processes that determine regional climate change, which span a wide range of spatial and temporal scales, and to the difficulty of extracting fine scale regional information from coarse resolution AOGCMs. Coupled AOGCMs are the modeling tools traditionally used for generating projections of climatic changes due to anthropogenic forcings. The horizontal atmospheric resolution of present day AOGCMs is still relatively coarse, order of 300-500 km, due to the centennial to millenial timescales associated with the ocean circulation and the computing requirements that these imply. However, regional climate is often affected by forcings and circulations that occur at the sub-AOGCM horizontal grid scale (e.g. Giorgi and Mearns 1991). As a result, AOGCMs cannot explicitly capture the fine scale structure that characterizes climatic variables in many regions of the world and that is needed for impact assessment studies (see Chapter 13). Therefore, a number of techniques have been developed with the goal of enhancing the regional information provided by coupled AOGCMs and providing fine scale climate information. here these are referred to as "regionalization" techniques and classify them into three categories: 1) high resolution and variable resolution ``time-slice" AGCM experiments; 2) nested limited area (or regional) climate models; 3) empirical/statistical and statistical/dynamical methods. Since the SAR report, a substantial development has been achieved in all these areas of research. This chapter has two fundamental objectives. The first is to assess whether the scientific community has been able to increase the confidence which can be placed in the projection of regional climate change caused by anthropogenic forcings since the SAR report. The second is to evaluate progress in regional climate research and to provide guidelines for the use of different methods. It is not the purpose of this chapter to provide actual regional climate change information for direct use in impact work, although the material discussed in this chapter serves most often for the formation of climate change scenarios. Climate scenario development is discussed in Chapter 13. Our assessment is based on an analysis of studies employing all the different modeling tools that are today available to obtain regional climate information. The analysis includes: a) an evaluation of the performance, strengths and weaknesses of different techniques in reproducing present day climate characteristics and in simulating processes of importance for regional climate; and b) an evaluation of the confidence and uncertainties in the simulation of climate change at the regional scale. In fact, even though a good simulation of present day climate does not necessarily imply a more accurate simulations of future climate change (see also Chapter 13), confidence in the realism of a model's response to an anomalous climate forcing can be expected to be higher when the model is capable of reproducing observed climate. Also, interpretation of the response is often facilitated by understanding the behaviour of the model in simulating the current climate. Based on this premise, the chapter is organized as follows. In the remainder of this section we present a summary of the conclusions reached in the SAR report concerning regional climate change and then briefly discuss in general terms the regional climate problem. In section 10.2 we examine the principles behind different approaches to the generation of regional climate information. Regional attributes of coupled AOGCM simulations are discussed in section 10.3. This discussion is important for two reasons: first, because AOGCMs are the starting point in the generation of regional climate change scenarios; and second, because many climate impact assessment studies still make use of output from coupled AOGCM experiments without utilizing any regionalization tool. Sections 10.4, 10.5 and 10.6 are devoted to the analysis of experiments using high resolution and variable resolution AGCMs, regional climate models and empirical/statistical and statistical/dynamical methods, respectively. In section 10.7 we then discuss studies in which different regionalization techniques have been compared, and in section 10.8 we summarize our main conclusions. 10.1.1 Summary of SAR The analysis of regional climate information in the SAR (section 6.6) consisted of two primary segments. In the first, results were analysed from an intercomparison of a number of coupled AOGCM experiments over 7 regions of the world. The intercomparison included coupled AOGCMs with and without ocean flux correction and focused on summer and winter precipitation and surface air temperature. Biases in the simulation of present day climate with respect to observations and sensitivities at time of CO$_2$ doubling were analyzed. A wide intermodel range of both biases and sensitivities was found, with marked inter-regional variability. Temperature biases were mostly in the range of +/- 5 C, with several instances of larger biases. Precipitation biases were mostly in the range of +/- 50%, but with a few instances of biases exceeding 100%. The range of sensitivities was lower for both variables. The second segment of the analysis mostly focused on results from nested regional models and downscaling experiments. Both these techniques were still at the early stages of their development and application, so that only a limited set of studies was available for the SAR. The primary conclusions from these studies were: a) both regional modeling and downscaling techniques showed a promising performance in reproducing the regional detail in surface climate characteristics as forced by topography, lake, coastlines and land use distributions; b) high resolution surface forcings significantly modify the surface climate change signal at the sub-AOGCM grid scale. Overall, the SAR still placed low confidence in the simulation of regional climate change produced by available modeling tools, primarily because of three factors: 1) Errors in the reproduction of present day regional climate characteristics; 2) wide inter-model variability in the simulated climatic changes; 3) sub-AOGCM grid scale structure of the climate change signal suggested by available regionalization studies. Other points raised in the SAR were the need of better datasets for model validation at the regional scale and the need to examine higher order climate statistics. \vskip .5cm \item{\it 10.1.2} {\it The regional climate problem} \vskip .3cm A definition of regional scale is difficult, as different definitions are often implied in different contexts. For example, definitions can be based on geographical, political or physiographic considerations, considerations of climate homogeneity, or considerations of model resolution. Because of this difficulty, in this chapter we adopt an operational definition based on the range of "regional scale" adopted in the available literature. From this perspective, we here define regional scale as describing the range of 10**4--10**7 km**2. The upper end of the range (10**7 km**2) is also often referred to as sub-continental scale. Circulations occurring at larger scales are clearly dominated by general circulation processes and interactions. Note that marked climatic inhomogeneity can occur within a region of 10**7 km**2 size in many areas of the globe. We refer to scales greater than 10**7 km**2 as ``large scale". The lower end of the range (10**4 km**2) is representative of the smallest scales resolved by current regional climate models. Scales smaller than 10**4 km**2 are here referred to as ``local scale". Given these definitions, the climate of a given region is determined by the interaction of forcings and circulations that occur at the large, regional and local spatial scales, and at a wide range of temporal scales, from sub-daily to multi-decadal. Large scale forcings regulate the general circulation of the global atmosphere. This in turn determines the sequence and characteristics of weather events and weather regimes which characterize the climate of a region. Embedded within the large scale circulation regimes, regional and local forcings and mesoscale circulations modulate the spatial and temporal structure of the regional climate signal, with an effect that can in turn influence large scale circulation features. Examples of regional and local scale forcings are those due to complex topography, land use characteristics, inland bodies of water, land-ocean contrasts, atmospheric aerosol, radiatively active gases, snow and sea ice distributions. Moreover, climatic variability of a region can be strongly influenced through teleconnection patterns originated by forcing anomalies in distant regions, such as in the El Nino Southern Oscillation (ENSO) and North Atlantic Oscillation (NAO) phenomena. The difficulty of simulating regional climate change is therefore evident. The effects of forcings at the global, regional and local scale need to be properly represented, along with the teleconnection effects of regional forcing anomalies. These processes are characterized by a range of temporal variability scales, and can be highly non-linear. Moreover, similarly to what happens for the global Earth system, climate at the regional scale is also modulated by interactions among different components of the climate system, such as the atmosphere, hydrosphere, cryosphere, biosphere and chemosphere. Therefore, a cross-disciplinary and multi-scale approach is necessary for a full understanding of regional climate change processes. This is based on the use of coupled AOGCMs to simulate the global climate system response to large scale forcings and the variability patterns associated with broad regional forcing anomalies. The information provided by the AOGCMs can then be enhanced via a suitable use of the regionalization techniques discussed in this chapter. --------------------------- 10.2 Deriving Regional Information: Principles, objectives and assumptions It is useful to present an overall discussion of the principles, objectives and assumptions underlying the different techniques today available for deriving regional climate change information. For some applications, the regional information provided by AOGCMs may suffice (10.2.1), while in other cases regionalization techniques can be used to enhance the regional information provided by coupled AOGCMs. The basic principles behind the regionalization methods we identified are discussed in sections 10.2.2, high resolution and variable resolution ``time slice" AGCM experiments; 10.2.3, regional climate models; and 10.2.4, empirical/statistical and statistical/dynamical models. The latter two techniques are often referred to as "downscaling" methods which use large-scale AOGCM information to derive consistent and detailed information at the regional and local scale. The concept of "downscaling" implies that the regional climate is conditioned but not completely determined by the large-scale state. In fact, regional states associated with similar large-scale states may vary substantially (e.g. Starr, 1942; Roebber and Bosart, 1998). The use of regionalization tools is advisable only when this enhances the information of AOGCMs at the regional and local scale. The "added value" provided by regionalization techniques depends on the spatial and temporal scales of interest as well as on the variable and climate statistics. This aspect of the regional climate problem is discussed in 10.2.5. Finally, the section closes with a brief overarching discussion of different sources of uncertaintiy present in the production of regional climate change information. 10.2.1 Coupled AOGCMs The majority of climate change impact studies have made use of raw climate information provided by transient runs with coupled AOGCMs without any further regionalization processing. The primary reason for this is twofold, i.e. the ready availability of this information, which is global in nature and is routinely stored by major laboratories, and the only recent development of regionalization techniques. Data can be easily drawn from the full range of currently available GCM experiments of the various modelling centres for any region of the World and uncertainty due to inter-model (or inter-run) differences can thus be evaluated (e.g. Hulme and Brown 1998). Also, data can be obtained for a large range of variables down to very short (sub-daily) time scales. In particular, spatially coherent climatic variability at short time scales is routinely simulated. >From the theoretical viewpoint, the major advantage of obtaining regional climate information directly from AOGCMs is the knowledge that internal physical consistency is maintained. The feedback resulting from climate change in a particular region on large scale climate and the climate of other regions is allowed for through physical and dynamical processes in the model. This may be an important consideration when the simulation of regional climate or climate change is compared across regions. The limitations of coupled AOGCM regional information are however well known. By definition, coupled AOGCMs cannot provide direct information at scales smaller than their resolution (order of several hundred km), neither can they capture the detailed effects of forcings acting at sub-grid scales (unless parameterized). Biases in the climate simulation at the AOGCM resolution can thus be introduced by the absence of subgrid scale variations in forcing. As an example, a narrow (subgrid scale) mountain range can be responsible for rainshadow effects at the broader scale. Many important aspects of the climate of a region (e.g. climatic means in areas of complex topography or extreme weather systems such as tropical cyclones) can only be directly simulated at much finer resolution than that of current AOGCMs. Analysis relevant to these aspects is undertaken with AOGCM output, but various qualifications need to be considered in the interpretation of the results. Past analyses have indicated that even at their smallest resolvable scales, which still fall under our definition of regional, coupled AOGCMs have substantial problems in reproducing present day climate characteristics. Many scientists maintain that the minimum skillful scale of a model is of several grid lengths, since these are necessary to describe the smallest wavelengths in the model and since numerical truncation errors are most severe for the smallest resolved spatial scales. Also, non-linear interactions are poorly represented for those scales closest to the truncation of a model because of the damping of dissipation terms and because only the contribution of larger scale (and not smaller scale) eddies is accounted for (e.g. von Storch, 1995). Advantages and disadvantages of using AOGCM information in impact studies can weigh-up differently depending on the region and variables of interest. For example, in instances for which sub-grid scale variation is weak (e.g. for mean sea level pressure) the practical advantages of using direct AOGCM data may predominate. Chapter 13 discusses the use of AOGCM information for climate change scenario development. Even if resolution factors limit the feasibility of using regional information from coupled AOGCM for impact work, coupled AOGCMs are the starting point of any regionalization technique presently used. Therefore, it is of utmost importance that coupled AOGCMs show a good performance in simulating circulation and climatic features that affect regional climates, such as jet streams and storm tracks. Indeed, most indications are that, in this regard, the performance of coupled AOGCMs is generally improving, because of both, increased resolution and improvements in the representation of physical processes (see chapter 8 of this report). 10.2.2 High resolution and variable resolution time-slice AGCM experiments Though simulations of many centuries are required to fully integrate the global climate system, for many applications regional information on climate or climate change is required for at most several decades. Over these time scales atmospheric GCM (AGCM) simulations are feasible at resolutions of the order of 100 km globally, or 50 km locally with variable resolution models. This suggests identifying periods of interest (or "time-slices") within AOGCM transient simulations and modeling these with a higher resolution or variable resolution AGCM to provide additional spatial detail (e.g. Bengtsson et. al., 1995; Cubasch et al., 1995). Such an AGCM can then be used to simulate the climate response to an anomalous forcing (e.g. changes in greenhouse gas (GHG) and sulphate aerosols) by direct inclusion of the forcing along with a consistent set of initial conditions and ocean surface boundary conditions. For the control simulation these could be derived from observations, an AOGCM control simulation or a transient AOGCM simulation using observed changes in atmospheric trace gases and aerosols. For the anomaly experiment there are several possibilities: Values could be used directly from an AOGCM anomaly simulation (equilibrium or transient) or derived from various perturbations of the control. The latter could take the form of an idealised uniform change in SSTs or spatially and seasonally varying changes derived from AOGCM simulations with matching radiative forcing changes. In a typical experiment (e.g. May 1999), two time slices, say 1961-1990 and 2071-2100, are selected from a transient AOGCM simulation. Time-dependent fields of SST and sea ice distribution are extracted and used as lower boundary conditions for a high resolution (or variable resolution) AGCM. Also, time-dependent GHG and aerosol concentrations (or aerosol forcing) in the AGCM experiments are the same as in the corresponding coupled AOGCM time slice. Initial atmospheric and land surface conditions for the AGCM experiments are also interpolated from the AOGCM fields. The philosophy behind the use of time-slice AGCM simulations is that, given the SST, sea-ice, trace gas and aerosol forcing, relatively high resolution information can be obtained globally or regionally without having to perform the whole transient simulation with high resolution models. The approach is based on two major assumptions. The first is that the large scale circulation patterns in the coarse and high resolution GCMs are not markedly different from each other, otherwise the consistency between the high resolution AGCM climate and the SST, sea ice and aerosol forcing from the coarse resolution AOGCM would be questionable. The other assumption is that the state of the atmosphere may be considered as being in equilibrium with its lower boundary conditions provided by the slower-evolving ocean and sea ice components. The main theoretical advantage of this approach is that the resulting simulations are globally consistent, capturing remote responses to the impact of higher resolution. Conversely this also allows the AGCM to evolve its own large scale climatology, possibly violating the first of the above assumptions. Thus it is important to consider the degree of convergence obtained at the standard and high resolutions. As resolution increases it is assumed that model simulations of the resolved large scale variables would asymptote to a solution. This implies there will be a threshold resolution greater than which the solution will not change fundamentally in character but just add extra detail at the finer scales. There is evidence that this has not been reached at the current resolution of AOGCMs which may add uncertainty to the value of regional information derived from AGCM timeslice experiments. A practical weakness of high resolution models is that they generally use the same formulations as at the coarse resolution at which they have been optimized to give accurate simulations of the current climate. The representation of some processes may thus be less accurate when finer scales are resolved so that some model formulations may need to be "re-tuned" for use at higher resolution. Experience with high resolution GCMs is still limited, so that presently increasing the resolution of an AGCM generally both enhances and degrades aspects of the simulations. With global variable resolution models this issue is further complicated as the model physics parameterizations have to be designed in a way that they can be valid and function correctly over the range of resolutions covered by the model. Another issue concerning the use of variable resolution models is that feedback effects from fine scales to large scales are represented only as generated by the region of interest. Conversely, in the real atmosphere feedbacks derive from different regions and interact with each other so that a variable resolution model based on a single high resolution region might give an improper description of fine-to-coarse scale feedbacks. In addition, a sufficient minimal resolution must be retained outside the high resolution area of interest in order to prevent a degradation of the simulation of the whole global system. Use of high resolution and variable resolution global models is computationally very demanding, which poses limits on the increase in resolution obtainable with this method. However, it has been suggested that high resolution AGCMs could be used to obtain forcing fields for higher resolution regional model experiments or statistical downscaling, thus effectively providing an intermediate step between coarse coupled AOGCMs and regional and empirical models. 10.2.3 Regional climate models What is commonly referred to as nested regional climate modeling technique consists of using output from global model simulations to provide initial conditions and time-dependent lateral meteorological conditions to drive high-resolution regional climate model (RCM) (or limited area model) simulations for selected time periods of the global model run (e.g. Dickinson et al. 1989; Giorgi 1990). SST, sea ice, GHG and aerosol forcing, as well as initial soil conditions, are also provided by the driving AOGCM. Some variations of this technique include forcing of the large scale component of the solution throughout the entire RCM domain (e.g. Kida et al. 1991; vonStorch et al. 1999) To date, this technique has been used only in one-way mode, i.e. with no feedback from the RCM simulation to the driving GCM. The basic strategy underlying this one-way nesting approach is that the GCM is used to simulate the response of the global circulation to large scale forcings and the RCM is used to account for sub-GCM grid scale forcings (e.g. complex topographical features and land cover inhomogeneity) in a physically-based way and to enhance the simulation of atmospheric circulations and climatic variables at fine spatial scales. The nested regional modeling technique essentially originated from numerical weather prediction, but is by now extensively used in a wide range of climate applications, going from paleoclimate to anthropogenic climate change studies. Over the last decade, regional climate models have proven to be flexible tools, capable of reaching high resolution (up to 10-20 km or less) and multi-decadal simulation times and capable of describing climate feedback mechanisms acting at the regional scale. A number of widely used limited area modeling systems have been adapted to, or developed for, climate application. On the other hand, the fundamental theoretical limitations of this technique are by now well known: lack of two-way interactions between global and regional climate; and effects of systematic errors in the driving large scale fields provided by global models. In addition, for each application careful consideration needs to be given to some aspects of model configuration, such as physics parameterizations, model domain size and resolution, technique for assimilation of large scale meteorological forcing (e.g. Giorgi and Mearns 1991, 1999). Recent studies have also shown that regional models exhibit internal variability due to non-linear internal dynamics not associated to the boundary forcing, which adds a further element to be considered in regional climate change simulations (Ji and Vernekar, 1997). Outstanding issues related to the above aspects of nested RCM modeling are discussed in section 10.5. >From the practical viewpoint, depending on the domain size and resolution, RCM simulations can be computationally demanding, which has limited the length of many experiments to date. An additional consideration is that in order to run an RCM experiment high frequency (e.g. 6-hourly) time dependent GCM fields are needed. These are not routinely stored because of the implied mass-storage requirements, so that careful coordination between global and regional modelers is needed to design nested RCM experiments. In this regard an aspect which should be considered for a specific demand of regional information is whether this can be obtained by simpler disaggregation methods. For instance, specification of topographically induced spatial detail in near-surface temperature may be possible with the use of GIS-based disaggregation schemes without having to rely on complex physical models (Agnew and Palutikof, 1999). Of particular interest is the direction taken by recent RCM modeling efforts towards the coupling of atmospheric models with other climate process models, such as hydrology, ocean, sea-ice, chemistry/aerosol and ecosystem models. The possibility of developing coupled "regional climate system models" will certainly open the use of RCMs to many new areas of global change research. 10.2.4 Empirical/statistical and statistical/dynamical downscaling Statistical downscaling is based on the view that regional climate may be thought of as being conditioned by two factors: the large scale climatic state, and regional/local physiographic features (e.g. topography, land-sea distribution and landuse; von Storch, 1995, 1999). >From this viewpoint, regional or local climate information is derived by first determining a statistical model which relates large-scale climate variables (or "predictors") to regional and local variables (or "predictands"). Then the large-scale output of an AOGCM simulation is fed into this statistical model to estimate the corresponding local and regional climate characteristics. A range of statistical downscaling models, from regressions to neural network and analogues, have been developed for regions where sufficiently good datasets are available for model calibration. In a particular type of statistical downscaling methods, called statistical-dynamical downscaling (see 10.6.3.3), use is made of atmospheric mesoscale models to develop the statistical models. Statistical downscaling techniques have their roots in synoptic climatology (Growetterlagen; e.g., Baur et al., 1944; Lamb 1972) and numerical weather prediction (Klein and Glahn, 1974), but they are also currently used for a wide range of climate applications, from historical reconstruction (e.g. Appenzeller et al., 1998, Luterbacher et al., 1999), to regional climate change problems (see section 10.6). A number of review papers have dealt with downscaling concepts, prospects and limitations: von Storch (1995), Hewitson and Crane (1996) and Wilby and Wigley (1998), Gyalistras et al. (1998), Murphy (1999a,b), Zorita (1999). One of the primary advantages of these techniques is that they are computationally inexpensive, and thus can be easily applied to output from different GCM experiments. Another advantage is that they can be used to provide local information, which can be most needed in many climate change impact applications. The applications of downscaling techniques vary widely with respect to regions, spatial and temporal scales, type of predictors and predictands, and climate statistics (from average temperature and precipitation to more episodic quantities such as storm interarrival times or frequency of strong wind events). The major theoretical weakness of statistical downscaling methods is that their basic assumption is often not verifiable, i.e. that the statistical relationships developed for present day climate also hold under the different forcing conditions of possible future climates. (TUNING SENTENCE) Another caveat is that these empirically based techniques cannot account for possible systematic changes in regional forcing conditions or feedback processes. The possibility of tailoring the statistical model to the requesetd regional or local information is a distinct advantage. However, it has the drawback that a systematice assessment of the uncertainty of this type of technique, as well as a comparison with other techniques, is difficult and may need to be carried out on a case-by-case basis. A number of examples are presented in section 10.6. An interesting by-product of empirical downscaling methods is that they offer a framework for testing the ability of physical models to simulate the empirically found links between large-scale and small-scale climate (Busuioc et al., 1999; Murphy, 1999a; Osborn et al., 1999; von Storch et al., 1993; Noguer, 1994). 10.2.5 The "Added Value" of regionalization techniques. The issue of "added value" of regionalization techniques is a difficult and much debated one. This is because it essentially depends on, and thus needs to be carefully formulated for, the specific scientific problem of interest. AOGCMs are designed to generate information at the large scale but, due to their resolution limitations, in many circumstances they are not expected to provide accurate regional and local climate detail. A fundamental question is therefore, whether it is possible to use regionalization techniques to add information about processes at the unresolved scales and their interaction with the climate system taking as input the large scale information from AOGCMs. The use of a regionalization tool for climate change simulation is thus adviceable to the extent that it produces additional information compared to the AOGCM. One of the reasons for developing regionalization techniques is to capture the effect of fine scale forcings in areas characterized by fine spatial variability of features such as topography and land surface conditions. In fact, in many regions topography and land use affect the spatial distribution of climate variables and generate (or modulate) atmospheric circulations at scales that are not explicitly described by AOGCMs. A regionalization method is thus needed to capture these effects and research has for example shown that the simulation of the spatial patterns of precipitation and temperature over complex terrain is generally improved with the increasing resolution obtained with regionalization techniques (see remainder of the chapter). The increased spatial resolution of regionalization tools also allows an improved description of regional and local atmospheric circulations such as synoptic and frontal extratropical systems, narrow jet cores, cyclogenetic processes, gravity waves, mesoscale convective systems, sea-breeze type circulations and extreme weather systems (e.g. tropical storms). Sub-grid scale processes that are parameterized in AOGCMs, such as cloud and precipitation formation, can also benefit from increased spatial resolution. Because spatial and temporal scales in atmospheric phenomena are often related, regionalization techniques can also be expected to improve the AOGCM information at high frequency temporal scales, such as daily or sub-daily. This is despite the fact that AOGCMs do provide high resolution temporal information. Therefore, for example, regionalization models can be used to improve the simulation of quantities such as daily precipitation frequency and intensity distributions, surface wind speed variability, storm inter-arrival times, monsoon front onset and transition times. >From a philosophical point of view, regionalization techniques are not intended to strongly modify the large scale circulations produced by the forcing AOGCMs, as this would result in inconsistencies between large scale forcing fields and high resolution simulated fields whose effects and implications would be difficult to evaluate. The assumption underlying this approach is that the effects of fine scale processes on the large scale fields is sufficiently well "parameterized" in the AOGCMs. In practice, the high resolution forcing described by some regionalization models, such as high resolution and variable resolution AGCMs and RCMs with sufficiently large domains, can yield significant modification of the large scale flows (e.g. storm tracks), possibly leading to an improved simulation of them. This has the important by-product of providing valuable information for the future development of higher resolution AOGCMs. 10.2.6 Uncertainties in the generation of regional climate change information There are several levels of uncertainty in the generation of regional climate change information. The first level, which is not dealt with in this chapter, is associated with emission and corresponding concentration scenarios (see Chapter 13). The second level of uncertainty is related to the simulation of the transient climate response by coupled AOGCMs for a given emission scenario. This uncertainty has a global aspect, related to the model global sensitivity to forcing, as well as a regional aspect, more tied to the model simulation of general circulation features. This uncertainty is important both, when AOGCM information is used for impact work without the intermediate step of a regionalization tool, and when AOGCM fields are used to drive a regionalization technique. The final level of uncertainty occurs when the AOGCM data are processed through a regionalization method. Sources of uncertainty in producing regional climate information are of different nature. On the modeling and statistical downscaling side, uncertainties are associated with imperfect knowledge and/or representation of physical processes, limitations due to the numerical approximation of the model's equations, simplifications and assumptions in the models and/or approaches, internal model variability, and inter-model or inter-method differences in the simulation of climate response to given forcings. It is also important to recognize that regional climate observations are sometimes characterized by a high level of uncertainty, especially in remote regions and in regions of complex topography. Finally, the internal variability of the global and regional climate system adds a further level of uncertainty in the evaluation of a climate change simulation. It is difficult to find unambiguous criteria to evaluate the level of confidence of a regional climate prediction, say for the 21st century, since this prediction is not directly verifiable. In general, a model's (or method's) capability of providing a good simulation of observed historical climate and climatic variability is an indication of increased confidence in the climate change simulation. Based on this criterion, a measure of uncertainty could be associated with the deviation of the model simulation from observed climate. This should however be viewed within the context that some model parameters are often optimized to reproduce present day climate and that, as mentioned, a good simulation of present day climate is not a sufficient condition for accurate simulation of climate change if the relevant processes and feedbacks that lead to the change are not well described. Another measure of confidence in the simulation of climate change is the model's ability to reproduce known climate conditions different from present, such as paleoclimates. A third measure of confidence can be related to the convergence of simulations by different models (or methods). Based on this criterion, a measure of uncertainty could be the spread of model (or method) results. Within this context, however, a convergence in model simulations might also indicate a commonality of basic flaws among models, since fundamental modeling assumptions are shared by most models. The emerging activity of seasonal to interannual climate forecasting may also give valuable insights into the capability of models to simulate climatic changes and may provide methodologies for evaluating the long term prediction performance of climate models. --------------------------- 10.3 Regional Attributes of A/OGCMs Since the SAR, GCM-based regional studies have been undertaken for North America (eg, Risbey and Stone, 1996, Fyfe and Flato 1999, Doherty and Mearns, 1999), South America (eg. Carril, Menendez and Nunez 1997, Labraga and Lopez 1997), Europe (eg. Barrow et al. 1996, Osborn et al. 1999, Raisanen 1998, Hulme and Brown 1998, Osborn and Hulme 1998, Benestad 1999), Australasia (e.g., Kidson and Watterson, 1995, Whetton et al. 1996a, Whetton et al. 1999), Southern Asia (e.g, Lal et al. 1998a,b, Dumenil 1998, Lal and Harasawa 1999a,b), Southeast Asia (e.g, Smith et al. 1998), Africa (eg, Joubert and Tyson, 1996, Fyfe 1998) and the South Pacific (Pittock, 1996, Jones et al. 1999). In addition, there have also been a number of studies which have considered multiple regions (e.g., Whetton et al 1996 b, Giorgi and Francisco 1999a,b Boer et al.1999a, Flato et al. 1999). These regional studies vary considerably with regard to: · Regional domains used. · Consideration given to within-region patterns. In some studies (e.g. Labraga and Lopez, 1997) this is a major focus. · Variables considered. Temperature and precipitation are most commonly considered, although often MSLP is considered as well (e.g. Schubert 1998). · The relative attention given to current climate validation as opposed to enhanced GHG changes. · Whether climatic variability and extremes are considered as well as climatic means. · The range of GCM experiments considered. This may be a single run only, a number of runs from a modeling center in which different forcing is used (e.g. GHG only versus GHG with sulphates - Boer et al. 1999a) and ensemble of runs with the same forcing and model (e.g. Hulme et al. 1999, Giorgi and Francisco 1999a), or runs with similar forcing from different models (e.g. Lal and Harasawa, 1999b, Giorgi and Francisco 1999b). · The age of the model runs. Some of the above studies include, amongst newer runs, some relatively old simulations. · Type of model run. Some use equilibrium 1xCO2 and 2xCO2 slab ocean GCMs (which are know to differ systematically from AOGCMs in their simulated pattern of regional climate change- Whetton et al. 1996a). This variation has allowed a range of relevant issues to be addressed and in the following sections studies such as those listed above will be drawn upon to illustrate various aspects of regional climate simulation with AGCMs. However, the large variety in study regions and methods also means that a comprehensive and consistent region by region analysis cannot be readily undertaken within the scope of this assessment. 10.3.1 Current Climate 10.3.1.1 Regional evaluation In this section we consider simulation of current regional climate using AOGCMs. Assessment of current climate simulation against observations (regional evaluation) is essential if the enhanced GHG regional results of models are to be correctly interpreted. Various issues associated with model evaluation at the global scale are discussed in Chapter 8 (section 8.2). In regional evaluation, the need to assess model errors arising due to the coarse horizontal resolution of current models is a major concern. Comparing model results against observations in regional studies can serve two objectives. The first is to assess the capability of the model to simulate the climatic features that are to be the focus of study under enhanced GHG conditions. Knowledge of this capability is important for the interpretation of the climate change results. For example, if the tropical low pressure systems simulated by models do not resemble tropical cyclones in key respects (e.g. insufficient intensity), the relevance of simulated changes in these systems for assessing changes in tropical cyclone behavior is problematic and needs careful interpretation. Many of the shortcomings revealed in comparisons of this sort will relate to the coarse horizontal resolution of the model, so that the regional evaluation can also be considered as an exercise in determining the skillful spatial scale of the model. The second objective is to assess how reliably the model simulates processes contributing to changes in key regional climatic features. Comparison of the climatic feature of interest against observations does contribute to this assessment, but there is a need to consider model climate more broadly. For example, if one's interest is in precipitation change, it would be appropriate to consider the model's simulation of synoptic circulation patterns associated with rainfall occurrence. Where regional precipitation is sensitive to sea-surface temperature patterns, it could be argued that regional oceanic processes require validation. Choosing appropriate variables for evaluation, determining the appropriate domain, weighing-up performance of one variable against another, and drawing overall conclusions on whether a model is performing acceptably well are difficult issues for which there is as yet no generally accepted methodology. 10.3.1.2 Climatic means Although current AOGCMs simulate well the observed global pattern of surface temperature (see Chapter 8), at the regional scale substantial biases are evident. To give an overview of the regional performance of current models, results are presented of Giorgi and Francisco (1999b) who compared model and observed seasonal mean temperature and precipitation averaged for each of the regions indicated in Figure 10.3.1. The AOGCM experiments they considered were a selection of those available through the IPCC Data Distribution Centre and included single simulations using the CSIRO, CCSR and MPI models, a three-member ensemble of CCC simulations and a four-member ensemble of HADCM2 simulations (see Table 9.4.1 for further model details). Figure 10.3.2 shows the biases in regionally averaged seasonal mean temperature and precipitation for 1961-1990 as analysed by Giorgi and Francisco (1999b). Temperature biases are typically within the range of +/- 3 K but exceed +/- 5 K in some regions, particularly in DJF. Precipitation biases are mostly between -20 and +50%, but exceed 100% in some regions, particularly in DJF. These regional biases are, in general terms, smaller than those of a similar analysis presented in the SAR (see also Kittel et al. 1998). For example, in the previous analysis regional temperature biases as high as 10-15 K were present in some models and regions. Given that the current analysis also includes many more regions, this difference in general performance strongly suggests that simulation of regional climate is significantly improved in current generation AOGCMs. Figure 10.3.3 illustrates model-to-model differences in the simulation of the annual cycle of regional precipitation and temperature for the examples of the Indian subcontinent and central Asia. These results are taken from the analysis of Lal and Harasawa (1999a) who also used a set of current AOGCMs runs available from the DDC (Table 9.4.1). Although in both regions all models generally reproduce the observed annual cycle of temperature, for some models and months errors as large as 5 K are present. The strong seasonal peak in precipitation over India from June to September is well simulated by all models, but the annual cycle in precipitation over Central Asia is poorly captured in a number of the simulations. Current generation AOGCM simulations in which historical changes in climate forcing over the 20th century are used enable simulated regional climatic trends to be assessed against observations. This was done by Boer et al. (1999a) for temperature and precipitation for the regions of southern Europe, North America, Southeast Asia, Sahel and Australia (defined as in the SAR). Simulated and observed regional linear temperature trends were in agreement for all regions except the Sahel for the runs where sulphate forcing was included. Little could be said about agreement in observed and model precipitation trends which were weak over the period in both the model and the observations. It should be stressed that assessments of model regional performance which are based on area-averaging of AOGCM output over large regularly-shaped regions (as was done in the studies reported above) should not be assumed to apply to all areas within these regions. Many of the regions considered contain a number of distinct climate regimes, and model performance may vary considerably from regime to regime. Climate change results may similarly differ. For the purposes of assessing model performance in a particular region, more detailed analysis is usually appropriate. Where studies have examined spatial patterns within regions (e.g. Joubert and Tyson, 1996; Labraga and Lopez, 1997), reasonable correspondence with observations were found, especially for temperature and MSLP. Most studies focus on seasonal mean conditions, but models can be analysed so as to focus on simulation of specific climate features. For example, Arritt and Goering (1999) examined circulation and precipitation patterns associated with the onset of the North American monsoon in simulations with the HADCM2 model and found this feature to be well simulated. It must be noted that some studies have identified important errors in current simulations of regional MSLP, such as the tendency for pressure to be too low over Europe and too high north and south of this area noted by Machenhauer et al. (1998). Such errors contribute significantly to local temperature and precipitation biases both in the global climate model and in nested high resolution RCM simulations (Risbey and Stone, 1996; Noguer et al. 1998, Machenhauer et al. 1998). As would be expected, GCM simulations of current climate are poorest at the local scale, particularly in areas of strong topographical controls (e.g. Schubert 1998). Widman and Bretherton (1999) concluded that means (and interannual variability) of precipitation can be well simulated in a mountainous areas only down to a resolution of three grid squares. However, in areas without complex topography, it is possible for the model results at individual gridpoints to compare well with observations, although it is necessary that the observations be averaged appropriately over the model grid boxes (Osborn et al. 1999). 10.3.1.3 Variability and extremes Interannual variability in temperature was assessed regionally, as well as globally, in a long control simulation with the HADCM2 model (Tett et al. 1997). Many aspects model variability compared well against observations, although there was a tendency for temperature variability to be too high over land. In the multi-regional study of Giorgi and Francisco (1999a), which used the ensemble HADCM2 simulations (Table 9.4.1), both regional temperature and precipitation variability were found to be overestimated. However, in a 200 year control simulation with the CCC model (Table 9.4.1), Flato et al (1999) noted that simulated interannual temperature and precipitation variability compared well with observations both globally and in five selected study regions (Sahel, North America, Australia, southern Europe and Southeast Asia). Validation of daily precipitation variability as simulated at gridboxes in GCMs is problematic because the corresponding variability in the real world operates at much finer spatial scale (e.g. Hennessy et al. 1997). A significant development in this area has been the work of Osborn and Hulme (1997) who devised a method of calculating grid box average observed daily precipitation data that corrected for biases commonly introduced due to insufficient station density. When they applied their method to output from the CSIRO GCM, daily precipitation performance was found to be better than that from a comparison based on simpler, and less appropriate, methods. Osborn and Hulme (1998) used this approach to assess the performance of a range of AGCMs in simulating rainfall variability over Europe. Models commonly simulated precipitation in winter to be more frequent and less intense than observed, although performance in some models was good in summer. Synoptic circulation variability at daily and longer time scales operates at a spatial scale which GCMs can simulate directly and there has been work focussed on GCM performance in this area at the regional scale (e.g. Huth, 1997, Joubert 1996, Katzfey and McInnes 1996, Osborn et al. 1999, Fyfe, 1998, Schubert 1998, Wilby et al. 1998a). Regions studied include North America, Europe, Southern Africa, Australia and East Asia. Although in many respects model performance is good, some studies have noted synoptic variability to be less than in the observations and the more extreme deviations from the mean flow to be less intense or less frequent than observed. For example, Osborn et al (1998) noted that HadCM2 output underestimated the frequency of the most intense flow situations over the UK. Finer scale circulation systems such as tropical cyclones can only be studied indirectly using GCMs (see Chapter 9). Osborn et al. (1999) also examined the relationship between the circulation anomalies and grid box average temperature and precipitation anomalies and found this to be well represented by the model. A similar investigation has been conducted by Busuioc et al. (1999) with regard to the representation of precipitation variability over Romania in versions of the ECHAM3 AGCM and model performance was found to be good in most seasons. Establishing the existence of appropriate links between variables in this way increases confidence in model performance under climate change conditions. 10.3.2 Future climate change 10.3.2.1 Climatic means With the exception of some restricted oceanic regions in some models, all regions of the globe show warming under enhanced GHG conditions. Typically the simulated regional warming is greater than the global average over land and over the northern higher latitudes in winter, and less than the global average in the tropics, the high southern latitudes, over the ocean, and in regions with strong local increase in sulphate forcing (see Chapter 9). Giorgi and Francisco (1999b) analysed regional temperature change in current AOGCMs under a range of forcing scenarios. In all regions warming depended strongly on the forcing scenario used. It was also noted that inter-model differences in simulated warming were large compared to differences between ensemble members from a single model. To considerable extent these inter-model differences would reflect differences in the global climate sensitivities of the models concerned. In some locations, additional regional factors can be identified as influencing the regional temperature response in a systematic way; for example Fyfe and Flato (1999) using the CCC model noted a tendency for greater warming at high elevations in the Rocky Mountains due to a snow-albedo feedback. To illustrate regional temperature change as simulated by current AOGCMs, Figure 10.3.4 presents some results from Giorgi and Francisco (1999b). The regions are as indicated in Figure 10.3.1, the range of simulations are from the same set of models as described for Figure 10.3.2 (including some ensembles), scenarios of 1% pa increasing CO2 with and without changes in sulphate aerosols are considered, and changes are for 2071-2100 compared to 1961-1990. In line with the globally averaged precipitation increase given by all models (see Chapter 9), precipitation is also simulated to increase regionally in the majority of cases. However, regions of precipitation decrease are also simulated. Figure 10.3.5 presents an analysis of simulated changes in regional precipitation equivalent to that for temperature presented in Figure 10.3.4. Note that although the figure includes some ensemble results, the following discussion will focus mainly on comparing the results of different models. Where CO2 only is increased (Figure 10.3.4a), most or all models show increased DJF precipitation for regions in the mid to high latitudes of the northern hemisphere. There is also some consistency on precipitation increase in the regions affected by the ITCZ at this time of the year. Simulated regional precipitation decreases in DJF are common in subtropical latitudes, but only for central America (CAM) and northern Australia (NAU) are decreases indicated by most or all models. The pattern is broadly similar in JJA, although with some features shifting northwards. Only the northernmost regions (ALA, GRL, NAS) show consistent increase, and simulated regional decreases are now common in the northern midlatitudes and the subtropics. Most models show decrease in the Mediterranean (MED) and central America (CAM) regions. Some regions along the ITCZ show increase (AMZ, SAS, SEA), but this is not true for the relevant African regions. The southern hemisphere midlatitude regions show inconsistent change (SSA, SAF) or decrease (NAU, SAU). When increased sulphates are included in the forcing scenario (Figure 10.3.4b) the results are similar, although there is some increase in frequency of simulated precipitation decrease in Africa (SAF, EAF) and southeast Asia (SAS, EAS, and SEA) in DJF. The magnitude of regional precipitation change varies considerably amongst models with the typical range being around zero to 40% where the direction of change is strongly indicated and around -20% to +20% where it is not. Larger ranges occur in some regions (e.g. -30% to +60% in southern Africa in JJA), but this occurs mainly in regions of low seasonal precipitation where the implied range in absolute terms would not be large. To illustrate further inter-model variations in simulated regional precipitation change we examine results obtained in model intercomparison studies for the Australian, Indian, North American and European regions. All of these regions have been extensively studied over the years using equilibrium 2xCO2 experiments (such as those featured in IPCC (1990)), first generation transient coupled AOGCMs (as in IPCC (1995) and more recent AOGCMs available in the DDC (Table 9.4.1). They are representative of a broad range of climatic regimes. This comparison also enables us to make some assessment of how the regional precipitation projections have changed as the models have evolved. In the Australian region, the pattern of simulated precipitation change in winter (JJA) has remained broadly similar across these three groups of experiments and consists of rainfall decrease in subtropical latitudes and rainfall increase south of 35-40 S (Whetton et al. 1996a, Whetton et al. 1999). However, as the latitude of the boundary between these two zones varied between models, southernmost parts of the Australia lay in the zone where the direction of precipitation change was inconsistent amongst models. The subtropical decreases ranged from around zero to minus 10% per degree of global warming with the larger decreases being more commonly found amongst the more recent coupled AOGCMs; the precipitation changes in the continental areas south of 35 S are typically in the range of plus or minus 5% per degree of global warming (Whetton et al. 1999). In summer (DJF) the equilibrium 2xCO2 experiments showed a strong tendency for precipitation increase over Australia, particularly in the northwest of the continent where changes as large +30% per degree of global warming were present in some models. This tendency was replaced in the first coupled AOGCMs by one of little change or precipitation decrease. This has remained the case when the most recent coupled models are considered. Whetton et al. (1996a) was able to partly attribute the contrast in the regional precipitation response of the two types of experiments to contrasts in their hemispheric patterns of warming. Differences in the precipitation change in the Australian region between simulations with and without sulphate forcing have not been thoroughly examined, although the results of Giorgi and Francisco (1999b) indicate little effect of sulfates. Together, Lal et al (1998b) and Lal and Harasawa (1999b) surveyed the results for the Indian subcontinent of seventeen climate change experiments including both equilibrium 2xCO2 and transient AOGCM simulations with and without sulfate aerosol forcing. In the simulations forced by GHG increases most models show wet season (JJA) rainfall increases over the region, although these increases are mostly less than 5% per degree of global warming. A minority of experiments show rainfall decreases. The experiments including sulphate forcing all showed reduced rainfall increases, or stronger rainfall decreases than their corresponding GHG only experiments. For North America we focus on the central Plains of the continent, which was established as one of the IPCC regions in the 1990 Report (Houghton et al., 1990). In the equilibrium 2xCO2 experiments reviewed in that report, there was a good deal of similarity of model response, with precipitation decreases prevailing in the summer and increases in the winter. Decreases and increases ranged within plus or minus 10%. In the second group of experiments (9 transient runs with AOGCMs) reviewed primarily in Houghton et al. (1996), a wider range of responses was found. In winter changes in precipitation ranged from about -12% to + 20% for the time of CO2 doubling, and most of the models (6 out of 9) exhibited increases. In summer the range of change was narrower, all within + and - 10%, but there was no clear majority response towards increases or decreases. Two of the most recent transient runs including aerosol forcing, the CCC AOGCM and the HADCM2 have been used in the first National United States Assessment Program, and were evaluated over North America by Doherty and Mearns (1999). The models simulated opposite changes in precipitation in both seasons. The CCC model simulated precipitation decreases (0.5 mm/day or 20%) in winter, and the HADCM2 increases of 0.5 mm/day, while in summer the CCC model simulated small decreases (-0.5 mm/day or 10%) and the HADCM2 mainly increases of the same magnitude. While all these results considered together indicate overall a tendency for more decreases to be simulated in the summer and more increases in the winter for the central US, there doesn't seem to be a striking reduction in the uncertainty for this region regarding changes in precipitation through the progression of climate models. For the European region, the evolution over the years of GCM simulations of enhanced GHG climate change has been examined by Hulme et al. (2000). They consider twenty-three climate change simulations (forced by CO2 change only) produced between the years 1983 and 1998 and including mixed-layer 1x and 2xCO2 equilibrium experiments as well as transient experiments. Figure 10.3.5 shows their results for simulated change in annual precipitation, averaged by latitude and normalised to percentage change per degree of global warming (to remove the effect of differences in forcing and model sensitivity). It may be seen that the consensus amongst current models for drying in southern Europe and wetter conditions in northern Europe represents a continuation of a pattern established amongst the earlier simulations. The effect of model development has primarily been to intensify this pattern of response. Variations from simulation to simulation in the regional enhanced GHG results of GCMs, which are particularly evident for precipitation, represents a major uncertainty in any assessment of regional climate change. Such variation may arise due to differences in forcing, systematic model to model differences in the regional response to a given forcing or differences due to natural decadal to inter-decadal scale variability in the models. Giorgi and Francisco (1999a,b) using the ensemble HADCM2 runs for various transient scenarios showed that uncertainty in forcing was clearly very important for regional temperature change, but less important, relative to other uncertainties, for precipitation change. Giorgi and Francisco (1999b) compared inter-ensemble differences in regional climate change (which can be viewed as representing the effect of the model's natural internal variability) with inter-model differences and concluded that the contribution of inter-ensemble differences to uncertainty in regional climate change was relatively small (this can be seen for precipitation change in Figure 10.3.4). However, it should be noted that Giorgi and Francisco (1999) used long (thirty year) means and large (sub-continental-scale) regions and that the uncertainty due to simulated natural variability would be larger when shorter averaging periods, or smaller regions, are used. The results of Hulme et al. (1999) also suggest that low frequency natural climatic variability is quite important at the subregional scale in Europe and can mask the enhanced GHG signal. Regional changes in the mean pattern of atmospheric circulation have been noted in various studies although typically the changes are not marked (e.g. Huth 1997, Schubert 1998). Indeed the work Conway (1998) and Wilby et al. (1998) suggests that the contribution of changes in synoptic circulation to regional climate change may be relatively small compared to that of non-synoptic processes. 10.3.2.2 Variability and extremes Giorgi and Francisco (1999a) found a tendency for interannual variability in regional precipitation to increase in HADCM2 under enhanced GHG conditions, but for temperature the response was less consistent across regions. Boer et al. (1999b) using the CCC AOGCM obtained marked decreases in interannual temperature variability over North America and Europe in DJF. They also noted a tendency for precipitation variability to increase. Beersma and Buishand (1999) analyzed monthly temperature and precipitation variability in a transient Hadley simulation over southern Europe, northern Europe and central North America. Between control and enhanced GHG samples of ten years duration few changes of statistical significance were found, although these included an increase in the standard deviation of precipitation of around 25% in winter, summer and autumn in northern Europe at the time of effective CO2 doubling. Furthermore they found that across the three regions and four seasons all substantial changes in precipitation variance were increases. It should also be noted that in many regions interannual climatic variability is strongly related to ENSO, and thus will be affected by changes in ENSO behavior (see Chapter 9). As noted in the SAR, many global climate models simulate increases in daily precipitation intensity and in the magnitude of extreme daily precipitation events. Kharin and Zwiers (1998) and Zwiers and Kharin (1999) have recently demonstrated this tendency at the global scale using the current version of the CCC GCM. Hennessy et al. (1997) addressed this topic regionally using both the CSIRO and Hadley centre models and found that under 2xCO2 conditions the one-year return period events in Europe, Australia, India and the USA increased in intensity by 10-25%. McGuffie et al. (1999) undertook similar analysis for the Sahel, North America, South Asia, Southern Europe and Australia using the BMRC model and various versions of the CCM model and also obtained decreased return periods for extreme precipitation events and increases in precipitation intensity. Daily temperature variability over Europe has been examined by Buishand and Beersma (1996) using the ECHAM/LSG model. They obtained statistically significant decreases of 15-30% around the time of CO2 tripling in the standard deviation of winter and spring temperature. Fewer studies have considered changes in variability and extremes of synoptic circulation under enhanced GHG conditions. Huth (1997) noted little change in synoptic circulation variability under equilibrium 2xCO2 conditions over North America and Europe. Katzfey and McInnes (1996) found that the intense cut-off lows off the Australian east coast became less common under equilibrium 2xCO2 conditions in the CSIRO model, although they had limited confidence in this result. 10.3.3 Summary Analysis of transient simulations with AOGCMs indicates that average climatic features are generally well simulated at the large and continental scale. At the regional scale, area-average biases in the simulation of present day climate are highly variable from region-to-region and among models. Temperature biases are typically within the range of +/- 3 K but exceed +/- 5 K in some regions, particularly in DJF. Precipitation biases are mostly between -20 and +50%, but exceed 100% in some regions. These regional biases are, in general terms, significantly smaller than those of a similar analysis presented in the SAR. Many aspects model variability compare well against observations, although there is a tendency for temperature variability to be too high over land. Simulated changes in mean climatic conditions for the late decades of the 21st century (compared to present day climate) vary substantially among models and among regions. All land regions undergo warming in all seasons, with the warming being generally more pronounced over cold climate regions and seasons. Average precipitation increases over most regions, especially in the cold season, as a result of an intensified hydrologic cycle. However, some exceptions occur in which most models concur in simulating decreases in precipitation. These include broad regions of Central America, Australia, Southern Africa and in the Mediterranean region in JJA. The magnitude of regional precipitation change varies considerably amongst models with the typical range being around zero to 40% where the direction of change is strongly indicated and around -20% to +20% where it is not. There is strong tendency for models to simulate regional increases in precipitation variability with associated reductions in the return period of extreme rainfall events. Increased interannual precipitation variability is also commonly simulated. However, changes in regional temperature variability vary considerably from model to model and region to region. --------------------------- 10.4 AGCMs with variable and increased horizontal resolution This section deals with the relatively new idea of deriving regional climate information from AGCMs with increased horizontal resolution. Although the basic methodology is suggested in the work of Bengtsson et al. (1995), where a high resolution GCM was used to predict changes in tropical cyclones in a warmer climate, it is only in the last few years that such models have been used more widely to predict regional aspects of climate change. Even so, only a limited number of experiments have been conducted to date and hence what follows is not a definitive evaluation of the technique but an initial exploration of its potential. For climate change applications, only AGCMs have been used to date. These consist of two versions of the Max Planck Institut’s ECHAM spectral AGCM, the Meteo-France/CNRM Arpege variable resolution spectral AGCM (run at both uniform high and variable resolution), two versions of the UKMO/Hadley Centre’s HadAM gridpoint AGCM, the LMD (Paris) variable resolution gridpoint AGCM and the MRI (Tokyo) JMA spectral AGCM (table 10.4.1). Instituti Model Horizonta Control Anomaly Region of on l Forcing Forcing interest Resolutio n MPI ECHAM3 T42 ECHAM/LSG ECHAM/LSG Euro/Glob al MPI ECHAM3/4 T106 Obs ECHAM/OPY Euro/Glob C al CNRM Arpege T213-T21 Obs/HadCM HadCM2 Euro/Glob 2 al UKMO HadAM2 0.83x1.25 Obs Global UKMO HadAM3 0.83x1.25 Obs HadCM3 Euro/Glob al LMD LMD-GCM 100-700km Obs Antarctic a MRI JMA T106 Obs MRI/GFDL/ Tropics +2K Table 10.4.1: High and variable resolution GCM control and anomaly simulations 10.4.1 Simulations of current climate Analysis of the current climate simulations of timeslice models has considered both deviations from the observed climate and effects on the model's climatology due to changes in resolution. Validation is generally performed on the sub- continental to global scale with high resolution information only considered for a particular region of interest. Most studies have considered just the mean climate and some measures of variability. The only extreme behaviour studied in any detail was the simulation of tropical cyclones. Even for mean climate, no comprehensive assessment of the surface climatology of variable or high resolution models has been attempted. Europe has been the most common area of study to date, although south Asia and Antarctica have also received attention. Thus, what follows is only indicative of the potential of the method and only raises selected concerns. 10.4.1.1 Seasonal mean climate The mean circulation is generally well simulated by AOGCMs, even though relatively large regional scale biases can still be present. Many features of the large scale climate are retained at higher resolution (Deque and Piedelievre, 1995, Stendel and Roeckner, 1998, Stratton, 1999a, May, 1999). A common deviation which has been found between ciarse and high resolution AGCMs is a poleward shift in the extra-tropical storm-track regions. It is suggested by Stratton (1999a) that this is linked to a general deepening of cyclones noted as a common feature in high resolution climate models by Machenhauer et al. (1996). More intense activity is also seen at higher resolution in the tropics, with a stronger Hadley circulation in ECHAM4 and HadAM3 that worsened the agreement with observations in both cases (Stendel and Roeckner, 1998 and Stratton, 1999b). The repositioning of the storm tracks generally improves the simulation in the northern hemisphere, resulting in a reduced positive polar surface pressure bias seen in the models at standard resolution. In the case of HadAM3, this leads to substantial improvements in northern hemisphere low level flow in winter (fig. 10.4.1). In the southern hemisphere, the impact is not consistently positive in all models, with the ECHAM and Arpege T106 simulations degrading features of circumpolar flow and HadAM3 and LMD models showins improvements (fig. 10.4.1, and Krinner et al., 1997). For surface pressure, in the tropics resolution appears to have little impact on the negative biases observed in both ECHAM4 and HadAM3. Increased resolution, however, improves the low-level south Asian monsoon flow in both models (Lal et al., 1997 and Stratton, 1999b). The fact that the above responses to increased resolution have common features amongst different models is an indication that they likely result from improved representation of the resolved variables. In contrast, an increase in the intensity of subtropical anti- cyclones observed in ECHAM4 results from a tropospheric warming promoted by excessive cirrus clouds. this was attributed to a scale-dependent response in the relevant parametrization (Stendel and Roeckner, 1998). AOGCMs generally perform worse in their simulations of surface climatology. This was one motivation for the first study employing a timeslice AGCM (Cubash et al., 1995), in which a T42 version of ECHAM3 was driven by the T21 ECHAM3/LSG AOGCM. The time-slice simulation provided a reasonable representation of the seasonal cycle of surface temperature over seven regions spanning the continents, but overall surface temperature was too high (by 2-5K), especially in summer. Precipitation was generally underestimated in summer, sometimes severely. Later experiments with the same model at T106 resolution (Cubasch et al., 1996) found that, over southern Europe, the winter temperature simulation did not improve with resolution and in summer the patterns improved but the positive biases became larger. As before, precipitation was still severely underestimated (by a factor of 2) in summer and the spatial precipitation patterns were improved at T106 resolution in summer but degraded in winter. Wild et al. (1995) showed that the summer warming in this simulation resulted from excessive insolation associated to reduced cloud cover. Updating the physics used for ECHAM4 improved some of these underlying biases but still the T106 simulation gave large negative precipitation biases and positive temperature biases in summer (Stendel and Roeckner, 1998). In simulations of European climate with the Arpege model, increasing resolution to T106 improved surface temperature simulation in both summer and winter due to reductions in the strength of a positive zonal flow bias. These biases were further reduced when using a stretched grid version T63s (stretching factor 3.5) with a resolution of T150 over most of Europe. Precipitation biases were also generally reduced with increasing resolution except for the too low values in south eastern Europe. The experience with HadAM2/3 was more mixed. As with Arpege, an improvement in the westerly flow at high resolution lead to improved temperatures and precipitation in winter throughout Europe and lso in summer over the north and west of Europe (Stratton, 1999 and Jones, 1999). However, a small warm and dry bias in the south east of Europe at standard resolution was greatly increased at higher resolution. This was caused by increased vertical activity at the higher resolution, promoting condensation and reducing a positive tropospheric humidity bias. This had the secondary impact of reducing cloud cover, increasing insolation, and reducing soil moisture in many areas of Europe to values which severely limited evapotranspiration (Jones, 1999). The same warming and drying in summer is seen over all extratropical continents in HadAM3 (Stratton, 1999) and clearly demonstrates a potential drawback of increasing the resolution of a model without comprehensively retuning the physics. Krinner et al. (1997) showed that to obtain a reasonable simulation of the surface climatology of the Antarctic with the LMD variable resolution AGCM many modifications to the model physics were required. With these modifications, the model was able to simulate surface temperatures to within 2-4K of observations and to provide a good simulation of the ice mass balance (snow accumulation), with both aspects being better than at standard resolution. 10.4.1.2 Variability and extreme events Many elements of the AGCMs flow-field intraseasonal variability, both intermediate frequency (or band-pass filtered) and low frequency, are simulated better at high resolution. Stendel and Roeckner (1998) show that in Echam4 low frequency variability of geopotential height and some eddy fluxes are simulated better at T106 than T42 (Gibson et al., 1997). In other cases, values underestimated at T42 are overestimated at T106. A similar picture in seen for HadAM2/3 (Stratton, 1999a,b) and in the Arpege model. In Arpege, at T106 there are also improvements in intermediate frequencies compared with the T63s. In contrast, Martin (1999) found little change with resolution in either the interannual or intraseasonal variability of circulation and precipitation of the south Asian monsoon in HadAM3. In studies with two high resolution AGCMs McDonald (1999) for HadAM3 and Yoshimura et al. (1999) for JMA have shown that they both produce realistic simulation of the location and frequency of tropical cyclones. In HadAM3 the frequency is somewhat overestimated in some areas but the annual cycle agrees well with observations. In addition McDonald (1999) demonstrates that whilst both standard and high resolution models capture large-scale features associated with tropical cyclones, the intensity and inner detail of cyclones is much more realistic at high resolution. 10.4.2 Responses to climate change Climate change studies with high resolution AGCMs have been limited to an even smaller selection of atmospheric models (table 10.4.1). The following results therefore cannot be regarded as fully representative, although some of the conclusions which are drawn from these studies point to important methodological strengths and weaknesses. 10.4.2.1 Applying anomalous atmosphere forcing When using a high resolution AGCM to simulate a climate change response consistent with that of an AOGCM experiment both the anomalous atmospheric forcing (GHG, sulphate aerosols etc.) and the accumulated effect of this on the ocean SST have to be provided as forcings to the AGCM. For the atmospheric forcing, GHG concentrations can be provided to the AGCM if their radiative effect is calculated by the model, or alternatively an equivalent amount of CO2 can be prescribed to give the same column integrated radiative forcing. For sulphate aerosols, if prescribed concentrations are used in the AOGCMs then these can be applied directly to the AGCMS. If the AOGCMs calculate the aerosol concentrations from prescribed sources then the AGCM can either use the same method or derive its concentrations from the AOGCMs. If the former method is used then the radiative forcing due to the aerosols may be different in response to possible scale dependencies in the sulphur cycle model or in the atmospheric circulations. If the latter method is employed and the large scale flow changes in the AGCm compared to the AOGCm, then the forcing will be inconsistent with the flow field. The most direct way of applying the oceanic surface forcing for an AGCM climate change experiment is to use AOGCM control values of SST and sea ice in the AGCM control experiment and anomaly AOGCM values in the AGCM anomaly experiment. This has the advantage of providing equivalent ocean surface forcing to the AGCM experiment. However, if there are substantial systematic errors in the control values, these could induce large biases in the atmospheric climatology of the AGCM. An alternative method is to use observed SSTs and sea-ice distribution for the control AGCM and then apply the changes in these values computed in the AOGCM anomaly experiment to provide anomalous AGCM values. This method provides forcing in the two model simulations which is consistent for SST changes, but not necessarily so for sea ice changes. 10.4.2.2 Changes in the mean climate The ECHAM3 timeslice climate change simulations reported in Cubash et al. 1996 predicted substantially different responses for southern Europe at T42 and T106 resolutions. For example, surface temperature in summer increased by over 4K over much of the region at T106 resolution, whereas at T42 resolution the response was generally less than +2K. Also, winter precipitation increased more in the T106 than the T42 experiments. In these cases, substantial differences in the control simulation were seen between the two resolutions, which would be an important factor in generating the response. For example, Wild et al. (1997) showed a large summer surface temperature positive bias at T106 resolution, implying a tendency towards soil drying which minimises soil moisture available for evaporative cooling in a warmer climate and hence enhances the response. Focusing over the whole of Europe, Deque et al. (1998) report a variable grid AGCM climate change experiment using SST forcing from HadCM2 control and GHG anomaly integrations. Their experiment predicts a moderate warming of maxima 2.5K in winter and 3.5K in summer over southern Europe and 1.5K and 1K, respectively, over northern Europe (fig. 10.4.2). In contrast, the driving AOGCM predicted greater warmings and a larger north-south gradient in winter (fig. 10.4.2). The reason for these differences result mainly from systematic errors in the control run Arpege large-scale flow, which is too zonal and too strong over mainland Europe. These errors, which are not present in HadCM2, enhance the moderating influence of the ocean SST. The precipitation responses show more similarities between the models, especially in summer, when both models predict a general decrease of up to 30% over most of Europe, a maximum decrease in southern Europe and small increases over north east Europe. However, the differences in the control simulations imply that confidence in this prediction is still low. In a similar experiment, HadAM3 was integrated at 1.25°x0.83° resolution with observed SSTs and sea-ice for the control and anomaly forcing from an HadCM3 GHG simulation. Globally, Johns (1999) found that in the annual mean at the largest scales many aspects of the timeslice response were similar to that in HadCM3. However, regionally or seasonally many differences are evident, notable examples being the land-sea contrasts and monsoon precipitation. Many circulation changes are also different in the two models, as could be expected given the differences in the control runs (see 10.4.1). In these experiments, the cloud feedbacks were found to be substantially different at the two resolutions, implying that changes in the parametrized processes are probably important in determining the responses. In contrast to this and the Arpege results, Jones (1999) showed many similarities in the large- scale patterns of the surface temperature and precipitation responses over Europe. Precipitation reduced everywhere in summer except the north west and increased almost everywhere in winter. Surface temperature increases during summer had a maximum in the south of Europe. The main difference in the response calculated by the two models is that surface warming in winter increases northwards in HadCM3 whereas it has a maximum east of the Baltic in the high resolution experiment. This is due to differences in the control simulations. HadCM3 predicts sea-ice which extends to the north Scandinavian coast in winter whereas in the HadAM3 simulation this area is ice-free. In the anomaly integration the ice-edge migrates polewards giving a much larger warming over this region in HadCM3. In this HadAM3/HadCM3 experiment the comparison of the responses is complicated by the use of observed sea-ice in the time slice control and model-derived sea ice changes in the anomaly run. A cleaner experimental design is used in an ECHAM4/OPYC simulations described by May (1999). In this experiment, the AGCM (ECHAM4) is run at T106 resolution and is driven by SSTs and sea-ice from a T42 AOGCM simulation. Two 30 year timeslices are simulated, 1970-99 and 2060-89, with with GHG and sulfate forcing from the IPCC IS92a scenario. The main inference from this study is that, as the future climate simulations are more similar to each other than the present day simulations, differences in the responses are due mainly to deviations in the control simulations. This conclusion largely rests on differences in extra-tropical circulations and precipitation over tropical land masses. Work currently in progress indicates that the positioning of the extra-tropical storm tracks in coarse and high resolution AGCMs is sensitive to the distribution of sea-ice. This may help explain the convergence of anomaly integrations noted above and help in the assessment of confidence in the responses in timeslice experiments. [DETAILS TO FOLLOW IN SECOND DRAFT] 10.4.2.3 Changes in variability and extremes In a T106 ECHAM3 simulation Bengtsson et al. (1996, 1997) found that under emhanced GHG conditions, the number of tropical cyclones decreased slightly in the Northern Hemisphere, and decreased by more than a factor of 2 in the Southern Hemisphere. This large difference in response for the two hemispheres raised questions about the model's ability to properly represent tropical cyclones at the resolution emplyed in the experiments. The tropical climate of ECHAM3 is quite sensitive to horizontal resolution, and methodological concerns were raised regarding the design of the experiment (Landsea 1997). Yoshimura et al. (1999) used a GCM of similar horizontal resolution to reexamine this issue. Under enhanced greenhouse conditions, they simulated a reduction in total tropical cyclone-like vortex formation in both hemispheres. This was despite the GCM simulation displaying an increase in rainfall in the tropics. Using a rather lower-resolution GCM (T42 NCAR CCM2), Tsutsui et al. (1999) built upon the previous work of Tsutsui and Kasahara (1996) to show basin-dependent changes in tropical cyclone formation under 2xCO2 conditions. Generally increased frequencies compared to the control climate were simulated in the western North Pacific, decreased frequencies in the North Atlantic, and similar frequencies in the southwest Pacific. These results agree with those of McDonald (1999) for the high resolution HadAM3 simulation, who also showed increases in tropical cyclones over the north Indian basin and a change in the timing of cyclones in the south-west Pacific. In the ECHAM3 simulation, Beersma et al. (1997) showed a general small decrease in north Atlantic cyclones, with regional increases in the North Sea. A decrease was found in the number of most intense depressions and an increases in the number of weak depressions. Again they questioned the significance of their results due to the small sample size. [DETAILS ON MAY AND ANDERSON WORK TO FOLLOW IN SECOND DRAFT] 10.4.3 Summary and recommendation Since the SAR variable and high resolution GCMs have been used more widely to provide high resolution simulations of climate change. Clearly the technique is still in its infancy with only a few modelling studies carried out and for only a limited number of regions. Also, there is little in depth analysis of the performance of the models and only preliminary conclusions may be drawn. Many aspects of the models' dynamics and large-scale flow are improved at higher resolution, though this is not uniformly so geographically or across models. Some models also demonstrate improvements in their surface climatologies at higher resolution. However, substantial underlying errors are often still present in high resolution versions of current AGCMs. Also, the direct use of high resolution versions of current AGCMs without some allowance of the dependence of models physical parametrizations on resolution leads to some deteriorations in the performance of the models. Changes in the large scale flow with increased resolution call into question the consistency between timeslice simulations and the SSTs and sea-ice forcings used to drive them. However, regional responses currently appear more sensitive to the AGCM than the SST forcing used. This result is partially due to some of the model responses being dependent on their control simulations and systematic errors within them. These factors and the small number of studies carried out imply that little confidence can be attached to any of the regional predictions provided by time slice simulations. The improvements seen with this technique are encouraging, but more effort should be put in analysing and possibly improving the performance of current models at high resolution. This is particularly important in view of the fact that future AOGCMs will likely use models approaching the resolution considered here in the next 5-10 years. --------------------------- 10.5 Regional Climate Models This section is mainly devoted to the use of nested RCMs to derive regional climate information from coarse resolution AOGCMs and to developments in regional climate modelling research since the SAR. The basic methodology (10.2.3) is inherited from numerical weather prediction with the pioneering use of RCMs due to Dickinson et al. (1989) and Giorgi (1990). 10.5.1 Overview of methodological developments, and improved understanding of weaknesses and strengths. Since the SAR many fundamental problems in the field of regional climate modelling have been studied (e.g. Giorgi and Mearns, 1999) although not all aspects of it have yet been fully explored. 10.5.1.1 Length of simulation. A fundamental motivation for the development of RCMs is to provide high-resolution information with a physical model over a limited area for time scales that would make a GCM simulation of comparable resolution prohibitively expensive. This goal, along with the relative simplicity of the nesting technique and the consistency it yields with AOGCM climate change simulations, explain the dramatic increase in the number of groups that have developed RCMs since the SAR. The generation of long term high-resolution RCM simulations, however, is still computationally demanding, both in terms of computer time and data storage. For this reason, to date much RCM work has focused on problems tractable with relatively short simulations (months to a few years). On the other hand, since the SAR it has been increasingly recognised that multi-year and possibly multi-decadal simulations should be used for climate change studies. This is for several reasons (e.g. Machenhauer et al. 1998): to increase the sample size and develop more meaningful climate statistics; to minimise problems related to internal model variability and variability of the climate system; to allow the model to fully reach internal equilibrium with the land surface conditions. While for the SAR only a few simulations longer than one year were available, several groups have now performed decadal simulations or longer, including one full transient experiment of 140 years in length (Hennessy et al. 1998). 10.5.1.2 The role of the lateral boundaries and domain size The influence of the lateral boundary forcing on RCM simulations when using variations of the standard Davies (1976) relaxation technique was extensively studied, for example, in the early work of Jones at al. (1995) and Cress et al. (1995). They showed that systematic errors in the driving fields from a GCM are generally transmitted to the nested RCM and numerous authors have confirmed this conclusion thereafter. Noguer et al. (1998) elaborated further on the influence of the external forcing by comparing 10 year RCM simulations driven by observed and GCM-derived boundary conditions. The simulation length allowed a decomposition of the systematic errors into internally generated and externally driven. Overall, 80-90% of the mean sea level pressure error was estimated to derive from the external forcing in all seasons. For mean surface temperature and precipitation the figures were lower, 40-60% in winter and 30-50% in spring and autumn. Errors in summer were mostly generated internally. In contrast, the study also found that the mesoscale signal generated by the RCM was relatively insensitive to the source of the boundary conditions. With the relative influence of boundary forcing and internal model physics potentially having a seasonal dependence, the choice of appropriate domain size for an experiment is not trivial. Should the latter dominate then the RCM solution may effectively decouple from the driving data. WIthin the context of downscaling, a climate change simulation exhibitng such an inconsistency can lead to problems in the intepretation of the results (Jones et al., 1997). Also, the domain size has to be large enough so that relevant local forcings and effects of enhanced resolution are not damped or contaminated by the application of the boundary conditions. Warner et al. (1997) suggested that if the area of meteorological interest has length scale L then the lateral boundaries should be at least a distance L/2 from this area. Seth and Giorgi (1998) showed that the effect of the location of the lateral boundaries could be especially important in studies of model sensitivity to internal parameters. Finally, a domain should be chosen so that its boundaries have minimal overlap with mountain ranges, since differences in resolution between driving and model fields may lead to inconsistencies and noise generation (e.g. Hong and Juan 1998). Contrary to the above experience, when choosing a domain for simulations of the Indian summer monsoon, Bhaskaran et al. (1996) showed minimal sensitivity of their results to the domain size. This was attributed to the main external forcing for the monsoon originating outside of all the model domains. The simulations also deviated little from the driving GCM indicating that the forcing was transferred to the RCM via the boundary relaxation technique without modification. To ensure full consistency between driving and nested model large scale fields Kida et al. (1991) and Sasaki et al. (1995) introduced the spectral nesting technique. Here the relevant components of the large scale driving fields force the low wave numbers of the RCM simulation throughout the entire model domain, while the RCM generates the higher frequencies. Spectral nesting, or nudging, has been further developed and refined in the recent works by Locke and Larow (1999), Waldron et al. (1996), McGregor et al. (1998) and von Storch et al. (1999). The spectral nudging ensures that the simulated model state remains close to the driving state at the large scales and, in this sense, can be viewed as an indirect assimilation technique. An alternate procedure that also ensures a close linkage of the RCM to the large- scale fields consists of frequent (daily or twice daily) restarts of the RCM based on large-scale initial driving conditions (Pan et al. 1999a). 10.5.1.3 Surface boundaries. It is by now well recognised that the surface forcing due to land, ocean and sea ice greatly affects a regional climate simulation. For example, Rinke and Dethloff (1999) found a substantial RCM sensitivity to sea-ice thickness and SST over the Arctic. In another study, Maslanik et al. (1999) illustrated that the sensitivity may derive from regions not resolved at the GCM grid scale. Similarly, land surface conditions significantly affected the RCM simulations of Pan et al. (1999b), Giorgi et al. (1996), Seth and Giorgi (1998), Christensen (1999), Pielke et al. (1999), and Chase et al. (1999). In particular, because most RCM experiments do not start with equilibrium conditions, initialisation of surface variables, such as soil moisture and temperature, is important. It is commonly assumed that a surface soil layer of a few cm depths reaches equilibrium after a few days or weeks. For the rooting zone (~ 1 m depth) the assumption is that soil equilibration is of the order of a few seasons, while for deeper soils the equilibrium time can be of years. Christensen (1999) gives an example where the time scale for soil temperatures to be in full balance with the atmosphere is longer than the characteristic diffusive thermal time scale of the soil layer. This resulted from non-linear interactions with the atmosphere. Unless the soil temperature was carefully initialised, it would not reach equilibrium for 2 to 4 years. Hence, incorrect soil temperature initialisation could produce a model drift which can significantly influence a temperature change signal. 10.5.1.4 Model resolution. To date, regional climate models have been mostly run at horizontal grid point spacing varying in the range of 20- 120 km, with a few experiments reaching grid point spacing of less than 10 km. In general, RCMs have shown a good performance in reproducing the effects of topographic and surface forcing at the selected resolution. In a study of the sensitivity of precipitation parameterisations to horizontal resolution in their RCM Giorgi and Marinucci (1996a) showed that the effects of physical forcings (e.g. topography) could be strongly modulated by the direct sensitivity of the model physics formulations to resolution. Laprise et al. (1998) also found substantially different behaviour of the same precipitation scheme in their RCM and the driving GCM. In a study over East Asia, Kato et al. (1999) showed that though the simulation of intense cyclonic events generally improved with increased resolution some aspects of model climatology deteriorated. A similar experience was reported by Christensen et al. (1998) in a double nested simulation of present-day climate at 18km resolution for Scandinavia. They found that the description of the hydrologic cycle improved with increasing resolution due to the better topographical representation but that some biases in the coarser resolution model where worsened. Within an RCM domain different sub-regions may require different resolutions to capture relevant forcings (e.g. topography). Double (one-way) nesting is one approach to achieve this objective. Another approach is to use two- way nested sub-grids, a capability available in many regional modelling systems but still not applied to regional climate problems. A third approach is to use smoothly varying horizontal resolution, or grid stretching, similar to that used in some global models, as described in the preliminary study of Qian et al. (1999). A subject which has received no attention in the published literature within the context of regional climate applications is that of the vertical resolution of RCMs. The number of vertical levels used in RCMs is generally between 10 and 30 and in many cases it is kept the same as in the driving GCMs. Increasing horizontal resolution has been shown to increase the variability and magnitudes of the vertical velocity (e.g. Jones et al. 1995) which suggests that, at least for stability reasons, vertical resolution should increase with horizontal resolution. 10.5.1.5 Model physics Traditionally, the development of regional climate models has followed two distinct approaches. In the first, a pre- existing (and well tested) limited area model system is suitably modified for climate application (e.g. in the model physics representations) and is used with driving conditions obtained either from analyses of observations or from different GCMs. This is the approach followed, for example, in the development of the NCAR RegCM (Giorgi et al. 1993a,b), the regional climate model of Miller and Kim (1996) and Kim et al. (1998), and the climate version of the CSU RAMS (Pielke et al. 1992, Copeland et al. 1996) and NCAR/PSU MM5 (e.g. Leung and Ghan 1999a,b). In the second approach, the full physics of a GCM is implemented within a regional dynamical framework, and the regional model thus obtained is mostly run using driving conditions from the host GCM. This approach is followed, for example, in the Canadian regional climate model (Laprise et al., 1988), the CSIRO DARLAM (McGregor and Walsh 1993), the MPI/DMI HIRHAM (Christensen et al. 1996) and the UKMO Unified model (Jones et al. 1995). These two approaches imply different strategies, which have both advantages and disadvantages. The strategy underlying the use of different physics parameterisations in the nested and driving models is that each set of parameterisations is developed and optimised for the respective model resolutions. The disadvantage of this strategy is that the interpretation of differences between nested model and driving GCM results is often difficult, because these may be caused not only by the different resolution forcings, but also by the differences in the physics schemes used. Another potential disadvantage is that the model physics schemes might result in such different forcings that spurious circulation can be produced in the interior of the domain. The strategy underlying the use of the same physics schemes in the nested and driving models is that maximum compatibility between the models is achieved. The main disadvantage with this approach is that physics schemes developed for coarse resolution GCMs may not be adequate for the high resolutions used in nested regional models. In addition, a parameterisation scheme (e.g. cumulus convection) can show a significant sensitivity to horizontal resolution (e.g. Giorgi and Marinucci 1996a, Laprise et al. 1998), and thus can present quite different behaviours in the nested and driving models. In some cases certain model parameters need to be re- calibrated for the particular resolution in order to give a satisfactory model behaviour (see also Section 10.4). Overall, both strategies have shown performance of similar quality (e.g. IPCC 1996), and depending on the particular experiment set up and model environment, either one may be preferable (Giorgi and Mearns 1999). In the context of climate change, if the physics schemes have similar behavior at coarse and fine resolutions, it may be preferable to use the same physics to provide consistency in the climate feedbacks associated with perturbations to the radiative forcing. 10.5.1.6 Coupling of atmospheric RCM with other components of the climate system. Several efforts have gone in the direction of coupling of atmospheric RCMs to other components of the climate system, such as ocean/sea ice, chemistry/aerosol, and land biosphere/hydrology models. An example of a coupled regional atmosphere/land/ocean/sea ice modelling system is ARCSyM; the Arctic Region Climate System Model originally developed by Lynch et al. (1995). ARCSyM has been used for a variety s studies of atmosphere/land/ocean interactions for the Arctic region (Lynch et al. 1997a,b, 1998; Bailey et al. 1997; Maslanik et al. 1999) and has been recently adapted to the Antarctic region (Bailey and Lynch 1999a,b). Weisse et al. (1999) coupled an RCM to a wave model and an ocean model to assess the role of an actual simulated sea state on the air-sea exchanges. Following the work of Hostetler et al. (1993), Small et al. (1999a,b) coupled atmospheric and lake models for the Aral Sea Basin, and their coupled model showed a remarkably good performance in reproducing the seasonal cycle of lake SST, sea ice extent, and the surface water and energy budgets of the lake. Leung et al. (1996) coupled their regional model (including a parameterisation of sub-grid scale topography and vegetation) to a basin-hydrology model, and were able to successfully simulate the hydrologic budget of basins characterised by complex topography. Miller and Kim (1996) and Kim et al. (1998) also carried out coupling between atmosphere and land hydrology models. Still, regarding biosphere-atmosphere coupling, Tsvetsinskaya et al. (1999a,b) coupled a crop model within the NCAR RegCM and then performed several experiments over the central Plains of the US to determine the effect of interactive seasonal plant growth on mesoscale patterns of temperature and precipitation. They found that the interactive model runs significantly affected surface fluxes and resulting tropospheric temperatures. Finally, RCM-aerosol interactive coupling ws first attempted by Qian and Giorgi (1999), who coupled the NCAR RegCM to a radiatively active aerosol source-transport-removal model, including both direct and indirect effects. they described various non-linear interactions between climate and aerosols. 10.5.1.7 Oceanic RCMs A large number of regional ocean models have been developed during the last decades for a wide variety of applications. However, the specific use of these models to climate change studies is very limited and only recent. In particular, Kauker (1998) used an approach similar to nested RCM modelling to develop a high-resolution ocean model for the North Sea. He completed multi-decadal present day and future ocean simulations driven by atmospheric forcing and lateral ocean forcing from an AOGCM experiment. While it is still early to evaluate the use of regional ocean models for climate change studies, and even though the resolution of some current global ocean models is already of the order of several tens of km, it is clear that the potential for this type of ocean model application is substantial and work in this direction should continue in the future. 10.5.2 Validation and simulations of present day climate Since the SAR, a vast number of RCM simulations have been conducted. McGregor (1997) presents an exhaustive review of simulations carried out until mid 1996, and many others have appeared in the literature since. It is not our intention to provide an exhaustive review of these experiments but rather to give an assessment of the general performance of RCMs in reproducing present day climate. An RCM can be validated by comparison with observations either for specific periods or for a long-term climatology. In the former case, observed (or "perfect") boundary conditions to drive the RCM are required, and are generally derived from NWP analyses or reanalyses (e.g. ERA, Gibson et al., 1997; or NCEP re-analysis, REF HERE?). Due to poor sampling in some areas and to observational uncertainty these are not error free. However, over most regions they will give accurate representation of the large-scale flow and tropospheric temperature structure. Multiannual RCM simulations with perfect driving boundary conditions can also be validated against long term climatologies. For GCM-driven experiments, in which the boundary conditions are obtained from GCM climate simulations, the caveats applied to GCM validation concerning the influence of sample size and decadal variability apply (see section 10.2, 10.3, and 10.4). Despite these caveats, relatively short simulations (several years) can identify major systematic RCM biases if they yield departures from observations greater than the observed natural variability (Christensen et al. 1997, Jones et al. 1999). Often a serious problem in RCM validation is the lack of good quality high-resolution observed data. While this data is available for some regions, over many areas of the globe observations are extremely sparse or not readily available. The Arctic and Antarctic regions are obvious examples but over many populated regions observation data sets either have low spatial resolution or are not easily accessible. Other examples of areas where observed data sets are problematic are regions characterised by complex terrain with insufficiently dense observing networks. In addition, only little work has been carried out on how to use point measurements to validate the grid-box mean values from a climate model, especially when using sparse station networks (e.g. Osborn and Hulme, 1997). A related issue is the type of data used for model evaluation. Most of the observational data available at typical RCM resolution (order of 50 km) is for precipitation and daily minimum and maximum temperature. While these fields have been shown to be useful for evaluating model performance, they are also the end product of a series of complex processes, so that the evaluation of individual model dynamical and physical processes is necessarily limited. Additional fields need to be examined in model evaluation to broaden the perspective on model performance and to help delineate sources of model error. Examples are the surface energy and water fluxes. Despite these problems, the situation is steadily improving (New et al. 1999a,b), with various groups developing high resolution regional observed climatologies (e.g. Frei et al. 1998, Christensen et al. 1998, VEMAP REFERENCE?). In addition, regional programs such as the Global Energy and Water Cycle Experiment (GEWEX) Continental-Scale International Program (GCIP) have been designed with the purpose of developing sets of observation data bases at the regional scale for model validation (GCIP, 1998). 10.5.2.1 Validation using simulations driven by analyses of observations Ideally, experiments using analyses of observations to drive the RCMs should precede any attempt to simulate climate change. The model behaviour in response to realistic forcing should be as close as possible to that of the real atmosphere and analyses of observation-driven experiments can reveal systematic model biases primarily due to the internal model dynamics and physics. A list of RCM simulations driven by analyses for one month or more and described in the literature is given in Appendix 10.A. For many of these a common measure of model skill is the regional bias of seasonally or monthly- averaged surface air temperature and precipitation, where the bias is defined as the difference betwee simulated and observed values. Table 10.5.1 presents regional biases of seasonally-averaged precipitation and surface air temperature for a sub-set of the experiments in Appendix 10.A in which a simulation of at least 3-year length was completed. This table indicates that current RCMs, when driven by analyses of observations, can reproduce observed average seasonal surface air temperature and precipitation over regions of size 10**5 -- 10**6 km2 with errors mostly in the range of +/- 0.5-2 K and +/- 5-40% (of observed precipitation), respectively. [CHECK THESE FIGURES BASED ON THE FINAL TABLE] In addition, various RCM intercomparison projects have been carried out to identify different or common model strengths and weaknesses. Christensen et al. (1997) compared 7 RCM simulations for summer and winter conditions over Europe using observed boundary conditions. The individual simulations used comparable resolutions (about 50 km) and included a common summer and winter month, though the domain sizes and length of simulation varied. A wide range of performance was reported, with the better models exhibiting a good simulation of surface air temperature except over southeastern Europe during summer. For winter precipitation, because of the strong forcing imposed by the boundary conditions, biases were derived mainly from errors due to the internal model physics and to a systematic tendency to simulate excessive cyclone activity. In summer, precipitation biases appeared to result from a partial decoupling of the RCM flow from the observed driving fields due to various deficiencies in the model physics. Tackle et al. (1999) presented results from the Project to Intercompare Regional Climate Simulations (PIRCS). In the first experiment, 7 models were compared in a simulation of the drought of summer 1988 over the continental U.S. Each model used a similar domain and resolution (60 km). A major finding was that the model ability to simulate precipitation episodes would vary depending on the scale of the relevant dynamical forcing. Organised synoptic-scale precipitation systems were generally simulated deterministically in that precipitation occurred at close to the same time and location as observed. Episodes of mesoscale and convective precipitation were represented in a more stochastic sense, with less degree of agreement with the observed events and among models both temporally and spatially. The performance of different models varied for different aspects of the simulation. An intercomparison of East Asian summer monsoon simulations from 3 models was presented by Leung et al. (1999). The primary result of this work was that cloud radiative processes in the models represented an important factor in determining differences between the model simulations and in determining model errors. The importance of cloud radiation processes in RCMs was further studied by Giorgi et al. (1999). An important development in RCM validation since the SAR is the extension of analyses from average climate to interannual variability. Studies in this direction were carried out by Luethi et al. (1996) for Europe, Giorgi et al. (1996) and Giorgi and Shields (1999) for the continental U.S., Sun et al. (1999b) for East Africa, Small et al. (1999a) for central Asia and Rinke et al. (1999) for an Arctic region. In all cases, the models were driven by ECMWF analyses. Overall the models showed a good performance in simulating interannual anomalies of precipitation and surface air temperature, both in sign and magnitude, over sub-regions of the domain varying in size from a few hundred to about 1000 km (Figure 10.5.1). These results indicate that, when driven by good quality large-scale fields, nested regional models can simulate well interannual surface climate variability at the sub- continental scale. The model performance can vary from region to region depending on the physiographic setting and the distance from the lateral boundaries. At the intra-seasonal scale, Fu et al. (1998) studied the evolution of the monsoon rain belt over East Asia for the period of April 1 to September 30, 1991. They demonstrated that the timing and positioning of the monsoon rain belt, as illustrated by a time-latitude cross-section of rainfall, was reproduced with a high degree of realism (Figure 10.5.2). A similar result was obtained by Emori et al. (1999) in an RCM study of the Baiu front over East Asia. Using the NCAR RegCM, Sun et al. (1999a) obtained a good simulation of intra-seasonal precipitation evolution over various regions of East Africa during the short rains of October - December 1988 (Figure 10.5.2). At even shorter time scales, Dai et al. (1999) examined the performance of the RCM of Giorgi and Shields (1999) in simulating the diurnal cycle of precipitation over the continental U.S. They showed that, despite good reproduction of climate averages and interannual variability over the region, the model still had significant problems in reproducing the observed diurnal cycles of precipitation, with the model performance varying substantially from region to region. 10.5.2.2 Simulations of present day climate using GCM boundary conditions Since the SAR, evaluation of RCMs driven by GCM simulations of current climate has gained much attention, and in fact many groups have performed GCM-driven experiments even prior to testing the models with analyses of observations. A list of GCM-driven regional simulation available in the literature is given in Appendix 10.B In general, the performance of RCMs in reproducing present day climate deteriorates when forced by GCM fields. Errors introduced by the GCM representation of large-scale circulation are trasmitted to the RCM as clearly shown by Noguer et al. (1998). However, since the SAR, regional biases of seasonal surface air temperature and precipitation have been reduced (Giorgi and Marinucci (1996b), Noguer et al. (1998) and Jones et al. (1999) for Europe, Giorgi et al. (1998) for the continental U.S. and Hennesy et al. (1998) for Australia). These improvements are due to both better large-scale boundary condition fields and improved aspects of internal physics and dynamics in the RCMs. Table 10.5.2 presents a summary of these and other representative results. TABLE 10.5.2 TO BE CONSTRUCTED AND INSERTED HERE Although the regionally averaged biases in the nested RCM are not necessarily smaller than those in the driving GCMs, all the experiments mentioned above, along with those of Leung et al. (1999a,b), Laprise et al. (1998), Christensen et al. (1998) and Machenhauer et al. (1998) clearly show that the spatial patterns produced by the nested RCMs are in better agreement with observations (Table 10.5.3). This is essentially due to the better representation of high-resolution topographical forcings and improved land/sea contrasts. [TABLE 10.5.4 2 WITH CORRELATION COEFFICIENTS: to be constructed] In most nested RCM experiments, by design, the average large-scale circulation in the nested and driving models are similar. However, when studying current climate this is not necessarily a constraint and Giorgi et al. (1998) showed that a nested RCMs could improve the large-scale circulations produced by the global model. This was especially so in summer when the effects of the internal model physics are most pronounced and lead to an improved large-scale precipitation patterns over the central U.S. This improvement was primarily attributed to a better representation of the Rocky Mountain chain in the RCM. Conversely, Machenhauer et al. (1998) showed that for three different present day climate simulations over Europe their nested model cyclones became too deep and traveled too far into the continent. They also demonstrated that interactions between the large-scale driving data and high resolution RCM forcings can have negative effects. During summer, the increased shelter due to the better-resolved mountains helped to enhance dry conditions over southeastern Europe. Some studies of additional climate variables have been performed. In a detailed study of the hydrologic cycle over Scandinavia, Christensen et al. (1998) showed that only at a very-high resolution do the mountain chains in Norway and Sweden become sufficiently well resolved to yield a realistic simulation of the annual evolution of the hydrologic cycle. (Figure 10.5.3). Confirming this result, Leung et al. (1999a) showed that only through the use of a sub- grid scale scheme capable of resolving complex topographical features a realistic simulation could be achieved of the seasonal evolution of snow formation and melting over the North-western U.S. (Figure 10.5.4). Noguer et al. (1998) used surface radiation observations from GEBA (Ohmura et al., 1989) elucidate the causes of warming in RCMs relative to their driving models. Only a few examples are available of analyses of variability in RCMs driven by GCM fields. At the intra-seasonal scale, Bhaskaran et al. (1998) showed that the leading mode of sub-seasonal variability of the south Asian monsoon, a 30-50 day oscillation associated with the northward migration of the circulation and precipitation anomalies, is more realistically captured by an RCM than in the driving GCM. Using the same model, Hassell and Jones (1999) showed that the RCM captured precipitation anomalies in the active and break phases of the monsoon (5-10 day periods of anomalous circulation and precipitation) that were absent from the driving GCM despite similar flow anomalies (Figure 10.5.5). At the daily time scales, Jones (1999) investigated the statistics of heavy precipitation events in RCM simulations. The RCM produced more realistic statistics of heavy precipitation events than the driving GCMs, capturing extreme events completely absent in the GCM. Much of this is due to the inherent disaggregation of grid-box mean values resulting from the RCM's higher horizontal resolution. However, even when aggregated to the GCM grid scale, the RCM wa closer to observations (as also demonstrated by Durman et al. 1999). 10.5.3 Climate change simulations Since the SAR several multi-year RCM simulations of anthropogenic climate change, either from equilibrium experiments or for time slices of transient simulations, have become available. These are given in Appendix 10.C. An important issue when analysing RCM simulations of climate change is the significance of the modelled responses. To date RCM simulations have been aimed at evaluating the models and processes rather than producing scenarios and have been relatively short (often only 5 years). At these timescales natural climate variability may mask all but the largest responses. In an analysis of RCM responses over Europe in four models, Machenhauer et al. (1998) concluded that generally only the full area averaged seasonal mean surface temperature and precipitation responses were statistically significant. In only a few cases across all seasons were subdomain deviations from the mean response significant. Also, in many of these cases the simulated climate responses were primarily attributed to a combination of systematic errors in the flow of the driving GCMs for present day conditions and to internal RCM model errors. Jones et al. (1997) estimated that at least a 30 year sample is required to confidently assess the mesoscale response in an RCM. Despite the limitations in simulation length, most RCM experiments clearly indicate that, while the large-scale patterns of surface climate change in the nested and driving models are similar, the mesoscale details of the simulated changes are quite different. Significantly different patterns of temperature and rainfall changes were found in the DARLAM 140 year long transient regional climate change simulation for Australia (Hennesy et al., 1998). This was most clearly seen in mountainous areas (Figure. 10.5.6). For example, winter rainfall in southern Victoria increased in the DARLAM simulation, but decreased in the driving GCM. Because of improvements in the DARLAM simulation of current climate relative to the GCM, they argued that its response was likely to be more plausible. A high resolution topographical modification of the regional precipitation change signal in a nested RCM simulation was also found by Jones et al. (1997) over Europe and Giorgi et al. (1998) over the continental U.S. All these studies illustrate the importance of fine resolution modelling of climate change in topographically complex regions. The response in an RCM can also be modified by changes in regional feedbacks. In a 20 year nested climate change experiment for the Indian monsoon region, Hassell and Jones (1999) showed changes in the regional warming patterns. A maximum of 5 K seen in central northern India in the GCM simulation was reduced and moved to north-west in the nested RCM, with a secondary maximum appearing to the south east (Figure 10.5.7). The shift of the main maximum was attributed to deficiencies in the GCM control climate that promoted excessive drying of the soil in northwest India. The secondary maximum was attributed to a complex response involving the RCM's better representation of the flow patterns in southern India resulting from an improved representation of the Western Ghats. In this instance, again it was argued that the improved realism of the RCM's control simulation increases confidence in its response. Changes in climate variability in control and doubled CO2 simulations for the US Great Plains are reported by Mearns (1999) and Mearns et al. (2000). They found significant decreases in daily temperature variability in winter and increases in temperature variability in summer. These changes were very similar to those of the driving GCM. Changes in variability of precipitation, however, were quite differentin the nested and driving models, particularly in summer, with increases being more pronounced in the regional model. In a 2CO2 regional climate scenario, Gallardo et al. (1999) also found that the Iberian Peninsula would be characterised by a higher seasonal variability than in the control. They report significant increase for surface temperatures (greatest in summer) and precipitation in winter. Nested RCMs can be used effectively for process studies. For example Giorgi et al. (1997) analysed the effect of doubling CO2 on the surface climate change signal over the European Alps. In their experiments, the simulated surface air temperature change signal due to CO2 conditions showed a marked elevation dependency, mostly during the winter and spring seasons, resulting in more pronounced warming at high elevations than low elevations (Figure 10.5.8). This was primarily caused by a depletion of the snow pack in doubled CO2 conditions and was enhanced by the snow-albedo-feedback mechanism. Interestingly, this result is consistent with observed temperature trends for anomalously warm winters over the alpine region (Giorgi et al. 1997). Changes in precipitation and other components of the surface energy and water budgets also showed an elevation signal. These results were confirmed by the RCM experiments of Leung and Ghan (1999b) and the GCM experiments of Fyfe and Flato (1999). In general, the presence of such elevation modulation of the climate change signal may have important consequences for climate change impacts on ecosystems and water resources in regions characterised by complex topographical systems. Another detailed study of particular climate change effects was carried out by Knutson et al. (1998), who analysed tropical hurricane intensities in the Northwest Pacific with the GFDL hurricane prediction system. Tropical storm-like features in a coarse mesh GCM control and climate change experiment were identified and corresponding 5-day simulations with a regional hurricane model were completed. Their main result was that the intensity of the hurricanes increased in a warmer climate because higher SST and increased environmental convective available energy (CAPE). Walsh and Ryan (1998) found a similar intensification of tropical cyclones near Australia using an RCM. Furthermore, Walsh and Katzfey (1998) identified a weak poleward shift of the tropical cyclone activity for double CO2 conditions using the same set of simulations. RCMs have also been used to explore the impact of land-use changes on regional climate by Pielke et al. (1999) and Chase et al. (1999). They found that the land-use changes due to human activities can induce climate modifications at the regional and local scale of magnitude similar to the observed climatic changes during the last century. The issue of regional climate modification by land0use change has been little explored within the context of the global change debate and, because of its potential importance, is in need of further examination. A simplified technique of using a RCM for climate change studies has been pioneered by Schar et al. (1996) and Frei et al. (1998). They forced a RCM with observed boundary conditions to simulate present day climate and developed a surrogate warmer climate by uniformly adding a temperature perturbation to the driving boundary conditions. Relative humidity was assumed to remain the same as present day, resulting in a domain-averaged 15% increase of the atmospheric moisture content. Their numerical experiments, carried out for Europe and for the autumn (i.e. the wettest season over the Alps), indicate a substantial shift towards more frequent events of heavy precipitation. The magnitude of the response increases with the intensity of the event and reaches several tens of percent for events exceeding 30 mm/day. Jones (1999) and Durman et al. (1999) found similar results using more rigorous multi-year GCM-driven simulations of control and doubled CO2 climate, e.g. 50-80% increases in events over 20 mm/day over parts of the UK. All these studies seem to indicate an increase in the frequency of high precipitation events in enhanced GHG climate conditions. 10.5.4 Summary and recommendation Since the SAR, significant improvements have been achieved in the areas of development and understanding of the nested regional climate modelling technique. These include many new RCM systems, multiple nesting, coupling with different components of the climate system (including aerosol-climate interactions) and research into the effects of domain size, resolution, boundary forcing and internal model variability. Nested RCMs have shown marked improvements in their ability to reproduce present day average climate, with much of this improvement due to better quality driving fields provided by GCMs. It is imperative for the effective use of RCMs in climate change work that the quality of GCM large scale driving fields continues to improve. New analyses have shown that RCMs can effectively reproduce interannual variability when driven by good quality fields. However, more analysis and improvements are needed of the model performance in simulating climate variability at short time scales (daily to sub-daily). Overall, the evidence is strong that regional models consistently improve the spatial detail of simulated climate compared to GCMs because of their better representation of sub-GCM grid scale forcings. This is not necessarily the case for region-averaged climate. The added value of the better topographic representation is especially relevant for the simulation of the surface hydrologic budget. Several RCM studies have been important to understand climate change processes, such as the elevation signature of the climate change signal or the effect of climate change on hurricanes. However, a consistent set of RCM simulations of climate change for different regions which can be used as likely climate change scenarios for impact work is still not available. Most RCM climate change simulations have been individual efforts aimed at specific goals. The need is there to coordinate RCM simulation efforts so that ensemble simulations with different models and scenarios for given regions can be developed to provide useful information for impact assessments. This will need to be achieved under the auspices of international or large national programs. Within this context an important issue is to provide RCM simulations of increasing length so as to minimize limitations due to sampling problems. Even if the best possible RCM could be considered to provide a best single scenario available for any region it would not be advisable to solely use this information in assessing regional climate change impacts. Different, but equally plausible, scenarios could be obtained by nesting the RCM in another set of GCM scenarios. WORK EXPECTED TO BE INCLUDED IN NEXT DRAFT: KATO ET AL. RCM RUNS OVER EAST ASIA --------------------------- 10.6 Empirical/statistical and statistical/dynamical methods 10.6.1 Introduction Formally, the concept of regional climate being conditioned by the large-scale state may be written as a stochastic and/or deterministic mapping of a predictor (a set of large-scale variables) on a predictand (a set of regional climate variables In general, the mapping is unknown and is modeled dynamically (i.e., through regional climate models) or empirically from observational (or modeled) data sets. In some cases the predictor and predictand are the same variables but on different spatial scales (for example the disaggregation schemes of Bürger, 1997; Wilks, 1999; and Widmann and Bretherton, 1999), but in most cases they are different. The mapping commonly employed is, in general, not designed to fully model all ranges of temporal scales. When using downscaling for assessing regional climate change, three implicit assumptions are made: - The predictors are variables of relevance and are realistically modeled by the GCM. Since different variables have different characteristic spatial scales, some variables are considered more realistically simulated by GCMs than others. For instance, derived variables (not fundamental to the GCM physics, but derived from the physics) such as precipitation are usually not considered as robust information at the regional and grid scale (e.g., Osborn and Hulme, 1997; Trigo and Palutikof, 1999). Conversely, tropospheric quantities like temperature or geopotential height are intrinsic parameters of the GCM physics and are more skillfully represented by GCMs. However, there is no consensus in the community about what level of spatial aggregation (in terms of number of grid cells) is required for the GCM to be considered skillful. For example Widmann and Bretherton (1999) find monthly precipitation on spatial scales of three grid lengths (in their case: 500 km) reliably simulated. - The transfer function is valid also under altered climatic conditions. This is an assumption that in principle can not be proven in advance. In the case of empirical functions, the observational record should cover a wide range of variations in the past; ideally, all expected future realizations of the predictors should be contained in the observational record. - Critical is the assumption that the predictors employed fully represent the climate change signal. Too little attention has in the past been paid to this assumption, but Hewitson (1999) and Charles et. al. (1999) have made progress in this respect. A diverse range of downscaling methods has been developed, but in principle fall into three categories, which are based upon the application of - weather generators, which are random number generators of realistically looking sequences conditioned upon the large-scale state (10.6.2.1). - transfer functions, where a direct quantitative relationship is derived through, for example, regression (10.6.2.2). - weather typing schemes based on the more traditional synoptic climatology concept (including analogs and phase space partitioning) and which relate a particular atmospheric state to a set of local climate variables (10.6.2.3). Each of these approaches has relative strength and weaknesses in representing the range of temporal variance of the local climate predictand. Consequently, the above approaches are often to some degree merged in order to compensate for the relative deficiencies in one method. Most downscaling applications have dealt with temperature and precipitation. However, a wide array of studies exists in which other variables have been investigated. Appendix XX provides a non-exhaustive list of past studies indicating predictands, geographic domain, and technique category. We are concentrating on references to applications since to 1995, since studies prior to that date made use of now outdated global climate change scenarios. [Appendix XX = Table 10.6.1] 10.6.2 Methodological options 10.6.2.1 Weather generators Weather generators are statistical models of observed sequences of weather variables. They can also be regarded as complex random number generators (Katz and Parlange, 1996), the outputs of which resemble daily weather data at a particular location (Wilks and Wilby, 1999). There are various types of daily weather generators, based on the approach to modeling daily precipitation occurrence: but the types usually fundamentally rely on stochastic processes. Two of these include the Markov chain approach (e.g., Richardson, 1981; Hughes et al., 1993, Lettenmaier, 1995; Hughes et al., 1999, Bellone et al., 1999) and the spell length approach (Racksko et al., 1991; Wilks, 1999a). In the spell length approach, which can be viewed as a natural way of extending the Markov chain approach, the length x of the spell lengths are simulated based on a probability distribution of the lengths. In the Markov chain approach precipitation occurrence is simulated day by day. Wilks (1999a) and Semenov et al. (1998) compare these methods. An additional approach is the so-called "conceptual model" approach, which involves chance mechanisms (e.g., clustering) by which storms arise and which is often used by hydrologists (O'Connell et al. 1999). Weather generators have been used for generating climate change scenarios that incorporate changes in climate variability (e.g., Katz, 1996; see Chapter 13, this volume) and for statistical downscaling, or for both simultaneously (Semenov and Barrow, 1997, Wilks 1999b). In the context of statistical downscaling the parameters of the weather generator are conditioned upon a large-scale state (see Katz and Parlange, 1996; Wilby et al., 1998; Charles et al., 1999), or relationships can be developed between large scale parameters sets of the weather generators and local scale parameters (Wilks, 1999b). Conditioning on large-scale states alleviates to some degree one of the chronic flaws of many weather generators, which is the underestimation of interannual variations of the weather variables (Wilks, 1989), and, to a degree, induces spatial correlation (Hughes and Guttorp, 1994). As is the case with other downscaling methods the success of this method is dependent upon the strength of the relationship between the stochastic generator parameters and the large-scale circulation index, and the stability of this relationship over time. As an illustration, the analysis of Katz and Parlange (1993, 1996) is discussed in some detail. They conditioned daily precipitation amount for a location in California on a circulation index, based on sea level pressure off the coast of California. They modeled the daily time series of precipitation as a chain dependent process, modeling occurrence as a first order Markov chain, and a power transform of intensity as normally distributed. The circulation index was allowed only two states, above and below normal pressure over a 78-year record. Using the Akaike and Bayesian Information Criteria they determined that model parameters such as mean intensity, standard deviation of intensity, and the probability of precipitation varied significantly with the circulation index state (high versus low pressure). They found that the conditioned model reproduced the precipitation variance statistics of the observations better than the unconditioned model, for example, interannual variance of monthly total precipitation. They went on to describe the use of their model for climate change scenario formation, i.e., conditions where the probability of obtaining a particular circulation index state is shifted. The mean precipitation changes linearly with the probability of the circulation state, but the standard deviation of the precipitation amount changes nonlinearly (Figure 10.6.1). These relationships indicate that the model allows for changes in the coefficient of variation of monthly total precipitation, which increases under mean drier conditions and decreases under mean wetter conditions. This method thus also allows for change in variability of precipitation along with the mean. 10.6.2.2 Transfer functions The more common approaches found in the literature are regression-like techniques or piecewise interpolations using a linear or nonlinear formulation. The simplest approach is to build multiple regression models relating free atmosphere grid cell values to surface variables. For example Sailor and Li (1999) have in this manner modeled local temperature at a series of US stations. Other regression models use fields of spatially distributed variables to specify local temperatures in Sweden (e.g.: Chen et. al., 1999), or principal components of regional geopotential height fields (e.g.: Hewitson, 1992). Canonical Correlation Analysis (e.g., von Storch and Zwiers, 1999) has found wide application. A variant of CCA is redundancy analysis, which is theoretically attractive as it maximizes the predictands variance; however, in practical terms it seems similar to CCA (WASA, 1998). Also Singular Value Decomposition has been used (Huth, 1999). Most applications have dealt with precipitation; for instance Busuioc and von Storch (1996) with Rumanian monthly precipitation amounts, or Dehn and Buma (1999) with a French Alpine site. Kaas et al. (1996) have successfully specified local pressure tendencies, as a proxy for local storminess, from large-scale monthly mean air pressure fields. Oceanic climate and climate impact variables have also been dealt with: salinity in the German Bight (Heyen and Dippner, 1998); and salinity and oxygen in the Baltic (Zorita and Laine, 1999); sea level (e.g., Cui at al., 1996); and a number of ecological variables such as abundances of species (e.g., Kroencke et al., 1998). In addition statistics of extreme events, expressed as percentiles within a month or season, have been modeled: storm surge levels (e.g., von Storch and Reichardt, 1997) and ocean wave heights (WASA, 1998). An alternative to linear regression is to use piecewise linear or nonlinear interpolation; geostatistics offers elegant "kriging" tools to this end (e.g., Wackernagel, 1995). The potential of this approach has been demonstrated by Biau et al. (1999), who related local precipitation to large-scale pressure distributions. Another approach is to use cubic splines, as was done by Buishand and Klein Tank (1996) for specifying precipitation in Switzerland. Also Hantel et al. (1998) adopt a nonlinear design for modelling snow cover duration in Austria with European mean temperature and altitude. Another non-linear approach is based on artificial neural networks (ANN; Hewitson and Crane, 1996), which are generally more powerful than other techniques, although the interpretation of the dynamical character of the relationships is less easy. For example, Trigo and Palutikof (1999) map with an ANN SLP and 500 hPa height values on daily temperature at a station in Portugal and find significantly improved specification as compared to a linear ANNs. Figure 10.6.2 shows two brief examples demonstrating two aspects of transfer function downscaling. The first involves transfer functions using predictors based on synoptic pattern (eg: through PCA,CCA techniques). As with weather typing approaches (see section 10.6.2.3 below), the use of pattern introduces a vulnerability to questions of stationarity under future climates. Such pattern dependence thus needs to be coupled with an analysis of the pattern stability under future climates. For example, Schubert (1998) utilizes PCA of synoptic fields with subsequent linear regression, and accompanies this with an analysis of the stationarity of synoptic pattern. Figure 10.6.2a demonstrates the skill of the transfer function downscaling in generatings the seasonal characteristics, derived from the dounscaled daily data. The second example demonstrates the advantage of transfer functions in preserving temporal evolution, and the common characteristic of a reduction of modeled variance. Cavazos and Hewitson (2000) use non-linear neural nets to derive transfer functions between GCM grid cell predictands (local to the target downscaling location) and daily precipitation. Figure 10.6.2b shows a time series of daily observed precipitation along with precipitation downscaled from the atmospheric predictors, demonstrating the skill in capturing the time-evolution of events and the reduction in variance. Figure 10.6.2.2b: From Cavazos and Hewitson (2000). Salamanca (Spain) observed daily precipitation (Po, total 102.9mm) and downscaled precipitation (Ps, total 92.9mm). Skill of 0.67, as measured by r2. 10.6.2.3 Weather typing This synoptic downscaling approach empirically defines weather classes related to local and regional climate variations. These weather classes may be defined synoptically or fitted specifically for downscaling purposes by constructing indices of airflow (Conway et al., 1996). The frequency distributions of local or regional climate are then derived by weighting the local climate states with the relative frequencies of the weather classes. Climate change is then estimated by determining the change of the frequency of weather classes. In many cases, the local and regional climate states are derived from the observational record. Wanner et al. (1997) used changing global to continental scale synoptic structures for understanding and reconstructing Alpine climate variations, while Widmann and Schaer (1997) could not relate changing Swiss precipitation to changing statistics of weather classes. Kidson and Watterson (1995) made a similar analysis for New Zealand. Jones and Davies (1999) apply the technique for studying changing air pollution mechanisms. The analog method was introduced into the downscaling context by Zorita et al (1995). Conceptually similar, but mathematically more demanding are techniques which partition the large-scale state phase space, for instance with Classification Tree Analysis, and use a randomized design for picking regional distributions. This technique was pioneered by Hughes et al (1993). Lettenmaier (1995) gives a general overview of these techniques. Both analog and CART approaches return the right level of variance and correct spatial correlation structures. In the following, we discuss in some more detail a case of statistical-dynamical downscaling as suggested first by (Frey-Buness et al., 1995): Statistical-dynamical downscaling (SDD) is a hybrid approach with statistical and dynamical elements. In a first step GCM results of a multi-year climate period are disaggregated into non-overlapping multi-day episodes of quasi-stationary large-scale flow patterns. Once defined, similar episodes are grouped in classes of different weather types. Typical members of these classes, i.e. episodes which in total comprise only a small fraction of the complete period, are simulated with a regional climate model (RCM). It is driven at its boundaries by the GCM results of the respective episodes. Eventually, the RCM results are statistically evaluated where the frequency of occurrence of the respective classes determines their statistical weight. An advantage over the SSD technique over other empirical downscaling techniques is that in this way spatially distributed local climates are specified. Its feasibility has been demonstrated by a series of studies on climate and climate change in the European Alps (see Appendix XX). As compared with conventional continuos RCM simulations (Section 10.5), the computational effort of SDD is small and almost independent of the length of the climate period. That this reduction of computational demands is not combined with a reduction is accuracy, at least in terms of time-mean distributions, is demonstrated by a comparison of mean precipitation distributions as simulated by a continuous RCM simulation and by the SSD technique. Figure 10.6.3 displays correlation coefficients and mean absolute differences, conditional upon the degree of disaggregation. When the computational load is reduced to 20%, the mean absolute error amounts to about 0.4 mm/day, whereas the correlation coefficient is about 0.96. Thus, in practical applications the intrinsic error of SDD is acceptable if the overall error is largely determined by the error of the used models (GCM and RCM). Figure 10.6.3. Similarity of time mean precipitation distributions obtained in a continuous RCM simulation and through SSD for different levels of disaggregation. Top: mean absolute difference [mm/day], bottom: spatial correlation coefficient. Horizontal axis: computational load of SSD. Ñ is the number of days simulated in SSD, N the number of days simulated win the continuous RCM simulation. 10.6.3 Issues in Statistical Downscaling 10.6.3.1 Temporal variance Transfer function approaches and some of the weather typing approaches suffer to varying degrees from an under-prediction of temporal climate variability, since only part of the regional and local temporal variability of a climate variable is related to large scale climate variations, while another part is generated regionally. (For the case of regression the mathematics of this principle are worked out by Katz and Parlange (1996).) Two approaches for bringing the downscaled climate variables to the right level of variability are in use: inflation and randomization. In the inflation approach, originally suggested by Karl et al. (1990), the variation is increased by the multiplication of a suitable factor; a more sophisticated approach, named "expanded downscaling", was developed by Bürger (1996). It is a variant of Canonical Correlation Analysis that ensures the right level of variability. This approach is utilized by Huth (1999) and Dehn et al. (1999). In the randomization approach the unrepresented variability is added as unconditional noise; that is, in the simplest case, the "missing" variance is added in the form of white noise, possibly conditioned on synoptic state (Hewitson, 1998). The concept is worked out in von Storch (2000), and applications are offered by Dehn and Buma (1999) and Buma and Dehn (1998). Conversely, weather generators suffer from the inverse of the above, and have difficulty in representing low frequency variance. However, conditioning the generator parameters on the large-scale state may alleviate this to some degree state (see Katz and Parlange, 1996; Wilks, 1999a; Wilby et al., 1998; Charles et al., 1999). 10.6.3.2 Validation The validation of downscaling techniques is an essential but difficult requirement. It requires demonstrating the robustness of the downscaling under future climates, and that the predictors used represent the climate change signal. Both assumptions are not possible to rigorously test, as no empirical knowledge is available so far. The analysis of historical developments as well as simulations with GCMs can provide support for these assumptions. However, the success of a statistical downscaling technique for representing present day conditions does not imply legitimacy for changed climate conditions (Charles et al., 1999). The classical validation approach is to specify the downscaling technique from a segment of available observational evidence and then assess the performance of the empirical model by comparing its predictions with independent observed values. This approach is particularly valuable when the observational record is long and documents significant changes in the course of time. An example is the analysis of absolute pressure tendencies in the North Atlantic by Kaas et al. (1996), who fitted a regression model which related spatial air pressure patterns to pressure tendency statistics. Similarly Wilks (1999) developed a downscaling function on dry years and found it functioning well in wet years. Hanssen-Bauer and Fĝrland (1998) and Hanssen-Bauer (1999) found in their analysis that data series of 50 year length may not be sufficient to derive a valid model. An alternative approach is to use a series of comparisons between models and transfer functions, as demonstrated by Busuioc et al (1999), Charles et al. (1999) and González-Ruoco et al. (199a,b). In the former study, it was first demonstrated that the GCM incorporated the empirical link; in the latter, a regional climate model was used. From these findings it was concluded that the dynamical models would correctly "know" about the empirical downscaling link; then the climatic change, associated with a doubling of carbon dioxide, was estimated through the empirical link and compared with the result of the dynamical model. In both cases, the dynamical response was found to be consistent with the empirical link, indicating the validity of the empirical approach and its legitimate approach in downscaling other global climate change information. 10.6.3.3: Choice of predictors The list of predictands in the literature is very broad and comprise direct climate variables (e.g.: precipitation, temperature, salinity, snow pack), monthly or yearly statistics of climate variables (distributions in wind speeds, wave heights, water levels, frequency of thunderstorm statistics), as well as impacted variables (e.g.: frequency of land slides). The Appendix XX lists a wide range of predictors, predictands, and techniques. Useful summaries of downscaling techniques and the predictors used are also presented in Rummukainen (1997), Wilby et al. (1998) and Wilby and Wigley (1999). However, outside of passing references in many studies to the effect that a range of predictors were evaluated, there is little systematic work that has explicitly evaluated the relevant skill of different atmospheric predictors (Winkler et al., 1997). The one commonality between most studies is, not surprisingly, the use of some indicator of the large-scale circulation. The choice of the predictor variables is of utmost importance. For example, Hewitson (1997, 1998) has demonstrated how the downscaled scenario of future change in precipitation may alter significantly depending on whether or not humidity is included as a predictor. The implication here is that while a predictor may or may not appear as the most significant when developing the downscaling function under present climates, the changes in that predictor under a future climate may be critical for determining the climate change. Some estimation procedures, for example stepwise regression, are not able to recognize this and exclude variables that may be vital for climate change. Such exclusion may lead to misleading scenarios of change. A similar issue exists with respect to downscaling temperature. Werner and von Storch (1993), Hanssen-Bauer (1999) and Mietus (1999) noted that low frequency changes in local temperature during the 20th century could not be related to changes in circulation. Schubert (1998) makes a vital point in noting that changes of local temperature under doubled atmospheric CO2 may not be driven by circulation changes alone, but may be dominated by changes in the radiative properties of the atmosphere. This is a particular vulnerability of any downscaling procedure in light of the propensity to use circulation predictors alone that do not necessarily reflect the changed radiative properties of the atmosphere. One possible solution is to incorporate the large-scale temperature field from the GCM as a surrogate indicator of the changed radiative properties of the atmosphere. This approach has been adopted by Dehn and Buma (1999) in their scenario of future Alpine land slides. Another solution is to use several large-scale predictors, such as gridded temperature and circulation fields (e.g., Gyalistras et al., 1998; Huth, 1999). After the availability of homogeneous re-analyses (Kalnay et al., 1996), the number of candidate predictor fields has been greatly enhanced (Solman and Nuñez, 1999); earlier, the empirical evidence about the co-variability of regional/local predictands and large-scale predictors was very limited and made many studies choose either gridded near surface temperature or air pressure, or both (Gyalistras et al., 1994). These "new" data sets will allow significant improvements in the design of empirical downscaling techniques. Taking advantage of these new data sets Cavazos and Hewitson (2000) systematically evaluate a broad range of possible predictors for daily precipitation. They conclude that, generalized across regions, the critical predictors are some indicator of mid-tropospheric circulation and humidity. Regions with a significant component of orographic rainfall also benefit from some indicator of surface flow. 10.6.4 Inter-comparison of downscaling methodologies An increasing number of studies comparing different downscaling studies have emerged since SAR. However, there is a paucity of systematic studies that use common data sets applied to different procedures over the same geographic region. A number of articles discussing different empirical and dynamical downscaling approaches (Giorgi and Mearns, 1991; Hewitson & Crane, 1996; Wilby and Wigley, 1997; Buishand and Brandsma, 1997; Rummukainen, 1997; Zorita and von Storch, 1997; Gyalistras et al., 1998; Kidson and Thompson, 1998, Murphy, 1999a,b, von Storch, 1999b, Biau et al., 1999) do present summaries of the relative merits and shortcomings of different procedures. These inter-comparisons vary widely with respect to predictors, predictands and measures of skill. A systematic, internationally coordinated inter-comparison project would be useful. The most systematic and comprehensive study so far is that one by Wilby et al. (1998) and Wilby and Wigley (1997). They compared empirical transfer functions, weather generators, and circulation classification schemes over the same geographical region using climate change simulations and observational data. The study considered a demanding task to downscale daily precipitation for six locations over North America, spanning arid, moist tropical, maritime, mid-latitude, and continental climate regimes. A suite of 14 measures of skill was used, strongly emphasizing daily statistics. These included such measures as wet spell length, dry spell length, 95th percentile values, wet-wet day probabilities, and several measures of standard deviation. Downscaling procedures in the study included two different weather generators, two variants of an ANN-based technique, and two stochastic/circulation classification schemes based on vorticity classes. The results prove to be illuminating, but require careful evaluation as they are more indicative of the relative merits and shortcoming of the different procedures, rather than a recommendation of one procedure over another. In the validation phase of the study the downscaling results were compared against the observational data, and indicated that the weather generator techniques were superior to the stochastic/circulation classification procedures, which in turn were superior to the ANNs. However, the superiority of the weather generator when validated against the observed data is misleading as the weather generators are constrained to match the original data (Wilby and Wigley, 1997). Similarly, the improved performance of the circulation classification techniques with regard to the ANNs is largely a reflection of the measures of skill used and indicates the tendency of ANNs to over-predict the frequency of trace rainfall days. In contrast, when the inter-annual attributes of monthly totals are examined the performance ranking of the techniques is approximately reversed with the weather generators performing especially poorly. The results indicate strength by weather generators to capture the wet-day occurrence and the amount distributions in the data, but less success at capturing the inter-annual variability (the low frequency component). The important question with this procedure is thus how to perturb the weather generator parameters under future climate conditions. At the other end of the spectrum the ANN procedures performed well at capturing the low frequency characteristics of the data, and showed less ability at representing the range of magnitudes of daily events.The stochastic/circulation typing schemes, being somewhat a combination of the principles underlying weather generators and ANNs, appear to be a better all-round performer. In application to GCM simulations of future climate, the procedures showed some consistency with the ANN indicating the largest changes in precipitation. However, assessing the relative significance of the changes is non-trivial, and at this level of inter-comparison the results of the climate change application are perhaps more useful in a diagnostic capacity of the GCM which appeared to show differences in the strength of the precipitation-circulation relationship. What is not evaluated in this study to any great degree is the range of variance spanned by each technique. Addressing this issue Wilby et al. (1998) and Conway et al. (1996) apply transfer functions to determine wet/dry probabilities and then use a stochastic procedure for the magnitude of precipitation, and in doing so capture some degree of the low frequency and high frequency variance. Zorita et al (1995) and later Cubasch et al. (1996) demonstrated that a suitably designed analog technique reproduces storm interarrival terms well. Similarly, Hewitson (1998) span the range of variance using an ANN transfer function to predict precipitation magnitude, and then stochastically model the residual variance as a function of atmospheric state. An additional factor not yet fully evaluated in any comparative study is that of the temporal evolution of daily events. In this respect the manner in which daily events develop may be critical in some areas of impacts analysis, for example hydrological modeling. While a downscaling procedure may correctly represent, for example, the number of rain days, the temporal sequencing of these may be as important. A number of analyses have dealt with the relative merits of non-linear and linear approaches. For example, Conway et al. (1996) and Brandsma and Buishand (1997) use circulation indicators as predictors and note that the relationships with precipitation on daily time scales are often non-linear. Similarly Corte-Real et al. (1995) effectively applied multivariate adaptive regression splines (MARS) to approximate non-linearity in the relationships between large-scale circulation and monthly mean precipitation. However, the application of MARS to large volume daily data may be more problematic (Corte-Real et. al., 1995). Other non-linear techniques are kriging and analogs, whose performance were compared by Biau et al., (1999) and von Storch (1999). Kriging resulted in better specifications of averaged quantities but too low variance, whereas analogs returned the right variance but lower correlations. Also analogs can be usefully constructed only on the basis of a large data set. It appears that downscaling of the short-term climate variance benefits significantly from the use of non-linear models. In particular, downscaling of daily precipitation benefits appreciably from the ability to better capture convective events. Most of the comparative studies mentioned above come to the conclusion that techniques differ in their success of specifying regional climate, and the relative merits and shortcomings emerge differently in different studies. This is not surprising, as there is considerable flexibility in setting up a downscaling procedure, and the suitability of a technique and the adaptation to the problem at hand varies. 10.6.5 Summary and Recommendations A broad range of statistical downscaling techniques has been developed in the past few years. Users of GCM based climate and climate change information may choose from a large variety of methods conditional upon their needs. Weather generators provide realistic sequences of events. With transfer functions statistics, like conditional means or quantiles, of regional and local climate may consistently be derived from GCM generated data. Techniques based on weather typing serve both purposes. Downscaling means post-processing GCM data; it can not account for insufficiencies in the driving GCM data. As statistical techniques are combining the existing empirical knowledge, statistical downscaling can describe only links which have been observed in the past. Thus, it is based on the assumption that presently found links will prevail under different climate conditions. It may be, in particular, that under present conditions some predictors appear less relevant, but become significant in describing climate change. It is recommended to test statistical downscaling methods by comparing their estimates with simulations with high-resolution dynamical models. The advent of decades-long homogeneous atmospheric re-analyses have offered the community many more atmospheric large-scale variables as possible predictors. Statistical downscaling requires the availability of long and homogeneous data series spanning the range of observed variance, while the computational resources needed are small. Therefore, statistical downscaling techniques are suitable tools for scientific communities without access to supercomputers and with little competence in process-based climate modeling. Often dynamical downscaling methods are providing much more information than may be needed in a specific application, so that resorting to the much simpler statistical techniques may often be advisable. Furthermore, statistical techniques may relate directly GCM derived data to impact relevant variables, like ecological variables or ocean wave heights, which are not simulated by contemporary climate models. It is concluded that statistical downscaling techniques is many cases a viable complement to process-based dynamical modeling, and will remain so in the future. --------------------------- 10.7 Intercomparison of methods Few formal comparative studies of different regionalization techniques have been carried out. To date, published work has mostly focused on the comparison between regional climate model and statistical downscaling techniques. Early applications of regional climate models for climate change simulations (Giorgi and Mearns, 1991; Giorgi et al., 1994) compared the models against observations or against the driving GCMs, but not against statistical/empirical techniques. Recently three studies have systematically evaluated a dynamical model against statistical/empirical techniques. Kidson and Thompson (1998) compared the RAMS dynamical model and a statistical regression-based technique. Both approaches were applied to downscale reanalysis data (ECMWF) over New Zealand to a grid resolution of 50 km. The statistical downscaling used a screening regression technique to predict local minimum and maximum daily temperature, and daily precipitation. The regression technique limits each regression equation to 5 predictors (selected from EOFs of 1000hPa and 500hPa geopotential height fields, local scalar wind speed and anomalies of geostrophic wind speed at 500hPa and 1000 hPa, anomalous 1000hPa-500hPa thickness and relative vorticity, and terms of vorticity advection). The results indicated little difference in skill between the two techniques, and Kidson and Thompson (1998) suggested that, subject to the assumption of statistical relationships remaining viable under a future climate, the computational requirements do not favor the use of the dynamical model. They also noted, however, that the dynamical model performed better with the convective components of precipitation. More recently Murphy (1998a) evaluated the UK Meteorological Office Unified Model (UM) regional configuration over Europe against a statistical downscaling model based on regression. A range of predictors similar to those used by Kidson and Thompson (1998) (EOFs and regional values of wind, vorticity, temperature, and additionally in this case, specific humidity) were used, with the difference that in this case monthly mean values were downscaled. The results showed similar levels of skill for the dynamical and statistical methods, in line with the Kidson and Thompson (1998) study. The statistical method was nominally better for summertime estimates of temperature, while the dynamical model gave better estimates of wintertime precipitation. Again the conclusion was made that the sophistication of the dynamical model shows little advantage over statistical techniques, at least for present day climates. Murphy (1998b) continued the comparative study by deriving climate change scenarios using GCM data from a coupled ocean-atmosphere future climate simulation (global configuration of the UM). Downscaling of the regional climates is from the same regional configuration of the UM, and the same statistical model. Unlike the validation study which compared the downscaling against observational data, the climate change situation showed larger differences between the statistical and dynamical techniques. The study concludes that the differences in the temperature downscaling do not derive from a breakdown of the statistical relationships, as might be suspected, but are perhaps related to different predictor/predictand relationships in the GCM. In contrast, the downscaled precipitation differences may stem from the exclusion of specific humidity in the regression equation, as moisture was a weak predictor of the natural variability. This point would seem to confirm the humidity issue raised in 10.6.3 (Hewitson, 1997, 1998; Hewitson and Crane, 1999). Mearns et al. (1999) compared regional model simulations and statistical downscaling using the RegCM regional model and a semi-empirical technique based on stochastic procedures conditioned on weather types which are classified from circulation fields (700hPa geopotential heights). While Mearns et al. suggest that the semi-empirical approach incorporates more physical meaning into the relationships, this approach does impose the assumption that the circulation patterns are robust into a future climate in addition to the normal assumption that the cross-scale relationships are stationary in time. For both techniques the driving fields were from the CSIRO AOGCM (Watterson et al., 1995). The variables of interest were maximum and minimum daily temperature and precipitation over central-northern USA (Nebraska). As with the preceding studies, the validation under present climate conditions indicated similar skill levels for the dynamical and statistical approaches, with some advantage by the statistical technique. Also in line with the Murphy (1999) study, larger differences were noted by Mearns et al. when climate change scenarios were produced. Notably for temperature, the statistical technique produced an amplified seasonal cycle compared to both the RegCM and CSIRO data, although similar changes in daily temperature variances were found in both the RegCM and the statistical technique (with the statistical approach producing mostly decreases). The spatial patterns of change showed greater variability in the RegCM compared to the statistical technique. Mearns et al. suggested that some of the differences found in the results were due to the climate change simulation exceeding the range of data used to develop the statistical model, while the decreases in variance were likely a true reflections of changes in the circulation controls. The precipitation results from Mearns et al. are in contrast to earlier studies with the RegCM producing few statistically significant changes (although both increases and decreases were indicated) and almost half the changes derived from the statistical technique (almost always an increase) being statistically significant. Overall, the above comparative studies indicate that for present climate both techniques have similar skill. Since statistical models are based on observed relationships between predictands and predictors, this result may represent a further validation of the performance of RCMs. Under future climate conditions more differences are found between the techniques, and the question arises as to which is "more correct". While the dynamical model should clearly provide a better physical basis for change, it is still unclear whether different regional models generate similar downscaled changes. With regard to statistical/empirical techniques it would seem that careful attention must be given to the choice of predictors, and that methodologies which internally select predictors based on explanatory power under present climates may exclude predictors important for determining change under future climate modes. --------------------------- 10.8 Summary and Conclusions Today a number of modeling tools are available to provide climate change information at the regional scale for impact assessment work. Coupled AOGCMs are the fundamental models used to simulate the climatic response to anthropogenic forcings and, to date, results from AOGCM simulations have provided the climate information for the vast majority of impact studies. On the other hand, resolution limitations pose severe constraints on the usefulness of AOGCM information, especially in regions characterized by complex physiographic settings. Therefore, in the last decade three classes of regionalization techniques have been developed to enhance the regional information of coupled AOGCMs: high resolution and variable resolution time-slice AGCM experiments, regional climate modeling, and empirical/statistical and statistical/dynamical approaches. Since the SAR substantial development has been achieved in all regionalization methods. It is important to stress that AOGCM information is the starting point for the application of all regionalization techniques, so that a foremost requirement in the simulation of regional climate change is that the AOGCMs simulate well the circulation features that affect regional climates. In this respect, indications are that the performance of current AOGCMs is generally improving. Analysis of different transient simulations with AOGCMs indicates that average climatic features are generally well simulated at the large and continental scale. Biases in the simulation of present day average surface climate variables are highly variable from region-to-region and among models. When looking at seasonal averages, regional biases in AOGCM simulations of present day climate are mostly (but not exclusively) in the range of +/- 3 K for surface air temperature and -20$\%$ to +50$\%$ of observed value for precipitation (with several instances of biases still exceeding 3 K for temperature and 100$\%$ for precipitation). Regional analysis of AOGCM transient simulations extending to 2100, for different scenarios of GHG increase and sulfate aerosol effects, and with a number of modeling systems (some simulations include ensembles of realizations) indicate that the average climatic changes for the late decades of the 21st century (compared to present day climate) vary substantially among models and among regions. The primary source of uncertainty in the simulated changes is associated with inter-model range of changes, with inter-scenario and intra-ensemble range of simulated changes being less pronounced. Despite the range of inter-model results some common patterns are emerging from AOGCM simulations of 21st century climate: 1) All land regions undergo warming in all seasons, with the warming (and inter-model range of results) being generally more pronounced over cold climate regions and seasons. This latter result is primarily related to the snow/sea ice albedo feedback mechanism and points to the importance of the description of cold climate processes in the models. 2) Average precipitation increases over most regions, especially in the cold season, as a result of an intensified hydrologic cycle. However, some exceptions occur in which most models concur in simulating decreases in precipitation. These include broad regions of Central America, Australia, Southern Africa and Southern South America in DJF and the Mediterranean region in JJA. Analysis in simulated interannual variability indicates that the AOGCM performance in reproducing observed variability varies across regions and models, but with a prevailing tendency for precipitation interannual variability (as measured by the standard deviation) to increase in future climate conditions. Various AGCMs have been used in time-slice mode and different variable resolution modeling efforts are under way. Although the number of available time-slice AGCM climate change experiments is still small, studies indicate that changing the model resolution also changes the model response to climatic forcings. In particular, the climate change response, as measured for example by the temperature change, is strongly dependent on the bias patterns in the present day simulations. The importance of using specific changes in forcing SSTs in time-slice AGCM experiments seemingly plays a secondary role. Since the SAR, a large number of RCM systems have been developed, with capability of high resolution multi-decadal simulations in a variety of regional settings. The analysis of RCM simulations has extended beyond simple averages to include higher order climate statistics, and has indicated that RCMs can reproduce well observed interannual variability given good quality forcing fields. More and improved high resolution climatologies for RCM validation have been developed since the SAR, but additional work is still needed in this regard, especially for remote regions and regions characterized by complex topography. Compared to AOGCMs, RCMs have been shown to improve the spatial patterns of surface climate as forced by topography and other sub-GCM scale processes. However, regionally averaged climate may not be necessarily improved. The increased resolution of RCMs also allows the simulation of a broader spectrum of weather events, in particular as concerning higher order climate statistics such as daily precipitation intensity distributions. Analysis of some RCM experiments indicate that this is in the direction of increased agreement with observations. The abundance of new regional model studies and the emerging coupling of RCMs with other components of the climate system illustrates the flexibility of regional models as tools for regional climate change research. A broad range of empirical/statistical and statistical/dynamical downscaling models are currently available which can be tailored to the specific needs of the user. These models have improved in particular since the advent of longer and better quality re-analyses of observations that can be used to develop the models. Empirical techniques can be easily implemented and applied to the output of different GCMs and do not require computationally intensive resources. Therefore they can be especially useful for groups and countries that do not have access to large computational resources. Measures of uncertaintt for statistical downscaling models are application-dependent and preliminary inter-comparison studies indicate that errors and uncertainties are of the same order of magnitude across methods and compared to physical models. Work performed with all these regionalization techniques indicates that substantial sub-GCM grid scale structure in the regional climate change signal occurs in response to regional and local forcings. This is because of the non-linear nature of the processes that regulate regional climate. In particular, modeling evidence clearly indicates that topography and the surface hydrologic cycle strongly affect the surface climate change signal. We conclude that the use of AOGCM information for impact studies needs to be taken cautiously, especially in regions characterized by pronounced sub-GCM grid scale variability in forcings, and that suitable regionalization techniques should be used to enhance the AOGCM results over these regions. A research area which has been little explored to date is that of regional effects of land-use change. The exploratory work of Pielke et al. (1999) and Chase et al. (1999) indicates that land-use change by human activities might produce local and regional changes in surface climate of similar magnitude as observed changes for the last decades. However, land use change has not been included in climate change experiments with AOGCMs and regionalization techniques. This issue clearly needs to be better addressed in future work. In principle, a simulation of regional climate change should consist of the following steps: 1) Developemnt of a range of emission and concentration scenarios; 2) Ensembles of coupled AOGCM simulations for each scenario with different models; 3) Use of different regionalization techniques, models and methods to enhance the regional AOGCM information. Considerations of various type may enter the choice of the regionalization technique, as different techniques may be most suitable for different applications and different working environments. High resolution AGCMs offer the primary advantage of global coverage and two-way interactions between regional and global climate. However, due to their computational cost, the resolution increase that can be expected from these models is limited. Variable resolution and RCMs yield a greater increase in resolution, with current RCMs reaching resolutions as fine as a few tens of km or even less. RCMs can capture physical processes and feedbacks occurring at the regional scale, but they do not represent regional-to-global climate feedbacks and they are affected by the errors of the AOGCM driving fields. Two-way nesting can capture regional-to-global feedback processes and some research efforts in that direction are currently under way. Statistical downscaling techniques offer the advantages of being computationally inexpensive, of providing local information which is needed in many impact applications, and of offering the possibility of being tailored to specific applications. However these techniques have limitations inherent in their empirical nature. The joint use of different techniques may provide the most suitable approach in many instances. For example, a high resolution AGCM simulation could represent an important intermediate step between coupled AOGCM information and RCM or statistical downscaling models. The convergence of results from different approaches applied to the same problem can increase the confidence in the results and differences between approaches can help to understand the behavior of the models. Despite recent improvements and developments, regionalization research is still a maturing process and the related uncertainties are still rather poorly known. One of the reasons for this is that most regionalization research activities have been carried out independently of each other and aimed at specific objectives. Therefore a coherent picture of regional climate change via available regionalization techniques cannot yet be drawn. More coordinated efforts are thus necessary to evaluate the different methodologies, intercompare methods and models and apply these methods to climate change research in a comprehensive strategy. ---------------------------