Many of the processes that govern precipitation occurrence and intensity occur at spatial and temporal scales that remain unresolved by the current generation of climate models. The resulting mismatch between the information reliably provided by such models and the needs of researchers assessing impacts associated with changes in precipitation has led to the development of downscaling techniques. In this paper, the need for downscaling is demonstrated by describing (1) the difference between real-world precipitation-governing processes and their representation in climate models and (2) the differences between theoretical expectations of precipitation changes at large and small scales in a warmer world. With a primary focus on empirical-statistical downscaling, the general strategy for developing regional scenarios is described along with outstanding issues that are likely to shape the next decade of downscaling research.
Climate scientists are increasingly confident that anthropogenic activities, primarily land use change and greenhouse gas emissions, are producing measurable effects on the global climate system (see, for example, the most recent report of the Intergovernmental Panel on Climate Change; IPCC 2007), including precipitation (Min et al. 2011; Pall et al. 2011). Recognition of this anthropogenic effect has fueled interest in the development of climate change scenarios driven by 21st century greenhouse gas trajectories, which are in turn derived from projections of population growth, economic factors, technological growth, climate change adaptation and mitigation, and other factors (Nakicenovic and Swart 2000; Moss et al. 2010).
Most commonly, atmosphere-ocean general circulation models (AOGCMs) are used to produce climate change scenarios consistent with the assumed greenhouse gas emission trajectory and resulting concentration pathway. AOGCMs are capable of reproducing the general characteristics of historical climate variability at large spatial (continental) and long temporal (seasonal) scales, but often fail to adequately reproduce observed climate statistics at smaller spatial and shorter temporal scales. For temperature, in the absence of terrain, this is not problematic since temperature tends to vary smoothly in time and space. Precipitation, on the other hand, is discontinuous in both space and time, making high resolution projections more challenging. This contribution will discuss the reasons for scale limitations in AOGCM precipitation scenarios and describe the basic methods that have been developed to overcome these limitations in the interest of developing local and regional scale precipitation scenarios at timescales relevant for assessment of climate change impacts.
The Precipitation Process
The processes that lead to precipitation are well understood by atmospheric scientists and are not expected to be different in a changing climate. Put simply, in moist air, adequate vertical motion leads to lower pressure and cooling of air which results in condensation of water vapor and formation of clouds. If droplets (ice crystals) grow sufficiently large, rain (snow) is initiated. Precipitation thus depends on both the supply of moisture (water vapor) and the presence of a lifting mechanism.
The amount of water vapor in air is primarily a function of large-scale temperature, but regional variations exist due to moisture transport and convergence, which is a function of the atmospheric circulation. The amount of water vapor needed to saturate air at a given temperature (usually termed the saturation vapor pressure) is described by the Clausius-Clapeyron equation. The equation can be found in most introductory atmospheric science texts (e.g., Stull 2000) and is presented graphically in Figure 1. If dew point temperature is substituted for air temperature, the Clausius-Clapeyron equation yields the actual vapor pressure. The exponential nature of the Clausius-Clapeyron equation dictates, for example, that as air warms, the amount of vapor needed for saturation becomes greater. The vertical motion needed for precipitation can result from several processes, including orographic lifting, convective instability, and baroclinic instability (i.e., midlatitude cyclones) and subsequent frontal lifting.
To understand how moisture availability and vertical motion govern precipitation, we first have to recognize that precipitation is a two-part process consisting of occurrence and intensity. Occurrence is a binary process; precipitation either occurs or does not occur. Intensity is zero if precipitation does not occur and takes on positive values if precipitation does occur. Both processes are governed by the combination of moisture availability and vertical motion. For example, dry air will need substantial vertical lifting to cool sufficiently for condensation. High humidity coupled with weak vertical motion may enhance heat stress, but will not lead to precipitation. When moisture content and vertical motion are sufficient to produce precipitation, the resulting precipitation intensity also depends on the amount of water vapor present (a function of not only the local humidity, but also the region of moisture convergence as described by Trenberth et al. 2003) and the intensity of the lifting. The characteristics of occurrence and intensity, along with their spatial and temporal variations, define the precipitation climate of a region as well as the hydrologic response to the precipitation climate. Clearly, the impact of the same precipitation total received over a different number of events can be substantial. Indeed, some climatological studies have suggested that it is the character of precipitation that is likely to change as opposed to the total amount (e.g., Trenberth et al. 2003).
As a first-order approximation of future precipitation change, we might consider how the two factors identified above, vertical motion and moisture availability, are likely to change under enhanced greenhouse gas forcing. Since orographic forcing is not likely to change on the timescales of interest (a few decades to perhaps a century), the primary driver of changes in vertical motion are likely to be circulation related, such as the position of the jet stream and subsequent cyclone tracks. In fact, this poleward shift in the polar jet stream and cyclone tracks has already been observed (McCabe et al. 2001; Wang et al. 2006; Archer and Caldeira 2008) and several modeling studies have projected further changes in the next century (Yin 2005; Bengtsson et al. 2006). The degree to which these changes have or will effect regional precipitation, as well as the attribution of the observed changes to anthropogenic greenhouse gas emissions is the subject of ongoing research.
Water vapor enters the atmosphere via the biosphere, land surface and the oceans through processes of evaporation and transpiration. These processes, and the resulting atmospheric water vapor burden, exhibit strong dependence on surface temperature. As temperature increases, evapotranspiration rates and atmospheric moisture content will also increase. Theoretical and modeling work suggests similar rates of change for air temperature and dew point temperature leading to roughly constant relative humidity at the global scale. With current globally averaged near-surface air temperature of approximately 287.5K (14.5 °C) and relative humidity of 77 percent, and a specified anticipated warming, we can use the Clausius-Clapeyron equation to determine the expected increase in atmospheric moisture content. This change is depicted as the arrow in Figure 1, based on a warming of 3 K. The change in humidity (depicted here as vapor pressure) is slightly greater than 7 percent per K of warming, in accord with the estimate of Trenberth et al. (2003). Trenberth et al. (2003) argue that rainfall intensity should also increase at approximately 7 percent per °C, although this number is considerably larger than the 1–2 percent per °C projected increase in precipitation described in the 4th Assessment Report of the Intergovernmental Panel on Climate Change (Meehl et al. 2007). The solution to this discrepancy is that, at the global scale, there must be a decrease in either the total frequency of precipitation or in the frequency of light to moderate precipitation events (Hennessey et al. 1997; Trenberth et al. 2003). Several recent studies have identified robust increases in heavy precipitation in agreement with these expectations (e.g., Groisman et al. 2005; Trenberth et al. 2007; Pryor et al. 2009).
Regional temperature changes associated with 21st Century global warming will exhibit considerable spatial variability leading ultimately to regional departures from the expectations of atmospheric moisture content based solely on global temperature change and the Clausius-Clapeyron relationship. When combined with regional changes in circulation, these regional warming variations strongly suggest that the constant relative humidity described in global analyses loses validity at the regional scale, as demonstrated by Wang et al. (2008) for North America. Development of regional precipitation projections therefore requires consideration of both regional warming and changes in circulation that might impact moisture convergence.
Simulation of Precipitation by AOGCMs
AOGCM development has paralleled computational development as the former depends critically on the latter. The complexity of climate models has therefore increased while their resolution (both spatial and temporal) has simultaneously improved. While there is confidence regarding AOGCM projections at continental and larger scales, the confidence is greater for temperature than precipitation (Randall et al. 2007) and there remains considerable uncertainty in regional precipitation changes derived from AOGCMs.
Precipitation is the product of complex interaction between cloud microphysics, convection, boundary layer processes, large-scale circulation, and others (Dai 2005). At small scales, AOGCMs are not capable of explicitly modeling convective processes, but instead use numerical parameterizations. Differences in these parameterizations are among the primary reasons for differences among AOGCMs outputs, as convection is also critical for vertical transfer of energy and water vapor among other processes (Tost et al. 2006).
The result of AOGCM shortcomings in modeling of precipitation processes leads to errors in (1) the ratio of convective to stratiform precipitation, (2) the diurnal cycle of precipitation (with most GCMs initiating daily convection too early), and (3) the frequency and intensity of precipitation (Dai 2006). For example, several studies (Ines and Hanson 2006; Sun et al. 2006; Schoof et al. 2009) have reported that AOGCMs produce precipitation far more often than observations but with too few heavy precipitation events. While the seasonal or annual totals may be similar to observed, the nature of the precipitation events is quite distinct. This is demonstrated for a single station in Figure 2, which shows the observed monthly precipitation frequency, wet-day intensity and total precipitation alongside values derived from AOGCM simulations. As shown, the models produce rain far too frequently, but with low intensity. The monthly totals are therefore sometimes correct, but for the wrong reasons. It is precisely the daily precipitation climatology that is critical for assessing climate change impacts related to precipitation, including those related to regional hydrology and agriculture.
The shortcomings of AOGCM precipitation simulations described in the previous section demonstrate the need for tools that bridge the gap between the information provided by AOGCMs and the data needed to drive regional climate change impact assessments. This process is known as “downscaling” and the techniques employed can be either dynamical or empirical-statistical.
Dynamical downscaling approaches utilize high resolution regional climate models (RCMs) with boundary conditions provided by the AOGCM. Operating at smaller spatial scales than their parent models, they include a greater range of processes and generally an improved representation of the land surface and other boundary features. Dynamical approaches have the advantage of physical consistency with their parent AOGCM. As computational capabilities have improved, larger dynamical downscaling experiments have become possible. For example, the recent North American Regional Climate Change Assessment Program (NARCCAP; Mearns et al. 2009) developed projections using approximately a dozen unique AOGCM-RCM combinations in a 50-km downscaling experiment for the conterminous United States.
While the dynamical modeling approach shows great promise, there are several drawbacks. Studies have shown that the choice of AOGCM and greenhouse gas concentration trajectory represent major sources of uncertainty in precipitation scenario development. The computation complexity of dynamical approaches precludes consideration of large ensembles of AOGCM simulations using multiple greenhouse gas scenarios. RCMs also tend to inherit bias from their parent AOGCM and most impact studies, particularly those in hydrology, have required statistical post-processing to correct bias (Fowler et al. 2007). Lastly, because meaningful information emerges on the scale of several grid points (around 100 km; Maraun et al. 2010), the finest RCMs cannot provide reliable information regarding precipitation at a point in space. Empirical-statistical downscaling represents an attempt to address these shortcomings by statistically modeling the link between large-scale climate and small-scale (often station-level) precipitation.
Empirical-statistical downscaling is grounded in the field of synoptic climatology, which views the near-surface climate as a function of both the large-scale circulation and the local environment (see Yarnal 1994). However, the approach has also been extended to include other large-scale variables (e.g., humidity) as predictors. The statistical approach is not without its drawbacks. First and foremost, the physical consistency noted for dynamical approaches does not pervade statistical approaches. Secondly, there is an assumption of stationarity (i.e., the relationship between large-scale climate and local/regional climate is stable through time and under changed climatic conditions) that cannot be tested for future time periods. If these shortcomings can be addressed by (1) choosing predictors and predictands on the basis of physical processes, and (2) at a minimum, testing the stationarity assumption using different segments of the historical records, the empirical-statistical approach can be a useful alternative and/or complement to dynamical approaches. A general depiction of the empirical-statistical downscaling process is shown in Figure 3. In the most common approach to empirical-statistical downscaling (termed Perfect Prognosis (PP)), a statistical relationship is established between the historical large-scale and regional- or local-scale climate. This relationship is then applied to AOGCM scenarios developed for another time period to investigate possible implications for surface climate. An alternative approach, Model Output Statistics (MOS) establishes the relationship directly between the historical AOGCM simulations and observed climate, thereby accounting for AOGCM bias.
Precipitation downscaling studies have been extremely diverse in terms of region and scope of application, choice of large-scale predictor(s) and regional- to local-scale predictands, and methodological approaches, which range in complexity from interpolation techniques to artificial neural networks, stochastic weather generators, and advanced pattern identification techniques (see recent reviews by Fowler et al. 2007 and Maraun et al. 2010). Despite a large number of publications focused on precipitation downscaling in the last decade, few universal rules regarding optimal predictors and methods have been identified. As noted by Fowler et al. (2007), few downscaling studies aimed at assessing hydrological impacts have adequately examined or addressed uncertainty or provided a probabilistic framework for developing scenarios. Consideration of multiple AOGCMs, greenhouse gas emissions trajectories, and downscaling approaches needed to gauge uncertainty increases the computational demands of the empirical-statistical downscaling approach substantially.
As a first step toward establishing a probabilistic downscaling framework for precipitation, Schoof et al. (2010) extended the probabilistic approach applied to wind downscaling by Pryor et al. (2005) to precipitation by focusing on the statistical processes that govern daily precipitation (i.e., the occurrence and intensity processes discussed in Section 2). The parameters of statistical models describing these processes were downscaled: The two parameters of a first-order Markov chain model for precipitation occurrence and the two parameters of the gamma distribution for non-zero precipitation amounts on wet days. This type of approach has several advantages, including (1) assessment of changes in wet or dry spells and (2) assessment of changes in different parts of the probability distribution (e.g., extremes). Separate consideration of occurrence and intensity also allows assessment of their individual contributions to changes in total precipitation as well as further investigation of the hypotheses regarding occurrence and intensity presented in Section 3. The results of Schoof et al. (2010) suggest that US regions are likely to differ with respect to changes in precipitation, with northern regions experiencing increases in winter precipitation (primarily due to greater intensity) and most regions experiencing drier summers (primarily due to less frequent occurrence). The study, however, used a single emission scenario and therefore cannot account for the full range of responses associated with changes in greenhouse gas forcing.
The findings of Schoof et al. (2010) and other regional precipitation projection studies reinforce the notion that global theoretical expectations for changes in precipitation do not translate to the regional scale. While some regions do experience slight decreases in occurrence but more extreme events described in Section 3, other regions experience different combinations of changes in occurrence and intensity with implications for seasonal and annual precipitation totals. The sources of these regional variations can be linked to circulation changes (i.e., changes in the intensity of midlatitude cyclones, poleward jet stream displacement, or changes in the intensity and location of semi-permanent circulation features such as subtropical highs) which ultimately affect moisture transport and convergence.
Summary and Conclusions
Anthropogenic greenhouse gas emissions are likely to impact precipitation climatology at different spatial scales. At large scales, theoretical arguments suggest that precipitation should increase as atmospheric moisture content increases in a warmer world. At the regional scale, there is greater uncertainty in precipitation projections due to likely changes in atmospheric circulation that will influence regions of moisture convergence. Such changes are notably difficult to quantify due to the coarse resolution of AOGCMs which represent the state-of-the-art in climate projection efforts. Downscaling methodologies have been developed to address the gap between information provided by AOGCMs and information needed to assess impacts in hydrology, agriculture, and other fields.
Uncertainty in the generation of regional precipitation scenarios has several sources including differences in AOGCM formulation, downscaling approaches, and a lack of knowledge regarding future greenhouse gas concentrations. Therefore, in order to fully investigate possible changes, downscaling needs to be conducted with multiple AOGCMs, multiple downscaling methods (both dynamical and statistical), and greenhouse gas concentration pathways, and preferably within a probabilistic framework that allows assessment of impacts across a range of sectors. Even then, the results may not be transferrable to other regions and may differ according to the impact assessed (e.g., impacts in hydrology vs. agriculture). This suggests that empirical-statistical downscaling research would benefit from a multi-model (statistical model in this case), multi-region approach that has been used in recent regional climate model transferability studies (e.g., Takle et al. 2007).
According to the recent review of Maraun et al (2010), there are several major challenges facing precipitation downscaling research. First, there are major shortcomings in data quantity and quality, especially in developing countries. Second, there is a difference in downscaling skill between precipitation from synoptic-scale systems and convective systems, which leads to greater uncertainty in projections for arid regions. Third, the diurnal cycle of precipitation is not well simulated in many dynamical models and few, if any, statistical models have been developed. Nevertheless, differences in the observed and modeled diurnal cycle have major implications for near surface energy budgets (Shin et al. 2007) and regional temperature responses (Pan et al. 2004). Fourth, most downscaling studies have focused on a single location, but many impacts, particularly those in hydrology, require spatial fields at resolutions similar to or smaller than RCM grids. A critical feature of such downscaled fields is spatial autocorrelation. Recent developments include spatial weather generators that can simulate multi-site series with appropriate inter-locational dependence (e.g., Wilks 1998; Khalili et al. 2009; Baigorria and Jones 2010), but such models have not yet been applied to assessing hydrological impacts associated with global climate change. Fifth, changes in small-scale processes will produce regional feedbacks that are not well captured by dynamical or statistical downscaling techniques. Lastly, downscaling methods will fail to provide useful projections of climate if parent AOGCMs have poor representations of climate variability and change.
Although AOGCMs have become increasingly skillful as computational capabilities have improved, they are likely to remain inadequate for assessing small-scale changes in climate and their associated impacts. This means that developments in climate downscaling will continue to contribute to our ability to assess the impacts of climate change at the regional scale for the foreseeable future. The ultimate utility of downscaling tools will depend on their ability to address the challenges outlined in this essay and thereby provide the best information to the impact assessment community.
Support from the National Science Foundation Geography and Regional Science Program (grants 0648025 and 0647868) is gratefully acknowledged. The views expressed are those of the authors and do not necessarily reflect those of the sponsoring agency. The AOGCM data used are from the data portal developed for the 4th Assessment Report of the Intergovernmental Panel on Climate Change. We gratefully acknowledge the international modeling groups for providing their data, the Program for Climate model Diagnosis and Intercomparison for collecting and archiving the model data, the JSC/CLIVAR Working Group on Coupled Modeling and the Coupled Model Intercomparison Project and Climate Simulation Panel for organizing the model data analysis activity, and the IPCC WGI TSU for technical support. The IPCC Data Archive at Lawrence Livermore National Laboratory is supported by the Office of Science, U.S. Department of Energy.
Author Bio and Contact Information
Justin Schoof is an Associate Professor of Geography and Environmental Resources at Southern Illinois University Carbondale (SIUC). He earned graduate degrees in Geography (Atmospheric Science Program) at Indiana University (M.S. 1999, Ph.D. 2004) and was a Post-doctoral Research Associate at the Center for Ocean-Atmospheric Prediction Studies (COAPS) at Florida State University before arriving at SIUC in 2006. He conducts research on statistical and applied climatology with a focus on regional climates. He may be reached at email@example.com or at the Department of Geography & Environmental Resources, Southern Illinois University, 1000 Faner Dr., Carbondale, IL 62901.