Search

This work presents a comprehensive assessment of the suitability of random forests, a well‐known machine learning technique, for the statistical downscaling of precipitation. Building on the experimental and validation framework proposed in the Experiment 1 of the COST action VALUE—the largest, most exhaustive intercomparison study of statistical downscaling methods to date—we introduce and thoroughly analyze a posteriori random forests (AP‐RFs), which use all the information contained in the leaves to reliably predict the shape and scale parameters of the gamma probability distribution of precipitation on wet days. Therefore, as opposed to traditional random forests, which typically provide deterministic predictions, our AP‐RFs allow realistic stochastic precipitation samples to be generated for wet days. Indeed, as compared to one particular implementation of a generalized linear model that exhibited an overall good performance in VALUE, our AP‐RFs yield better distributional similarity with observations without loss of predictive power. Noteworthy, the new methodology proposed in this paper has substantial potential for hydrologists and other impact communities which are in need of local‐scale, reliable stochastic climate information. Plain Language Summary: Statistical downscaling methods aim to improve the limited spatial resolution of current climate models by linking a set of key large‐scale predictor variables (e.g., geopotential, winds, etc.) to the predictand of interest (e.g., precipitation). Recently, the Experiment 1 of the COST action VALUE carried out the most comprehensive intercomparison of statistical downscaling methods to date. However, it lacked the inclusion of machine learning techniques, whose popularity has rapidly grown during the last years. Therefore, building on the same data and experimental framework used in VALUE, this work aims to partially fill this knowledge gap by introducing a modification of random forests—a well‐known machine learning technique—for stochastic downscaling of precipitation at 86 European locations. As opposed to traditional random forests, which typically provide deterministic predictions, our proposed model predicts a probability distribution of precipitation for each predictors' state. This is key to appropriately characterize the uncertainty of the downscaled predictions, allowing us to produce realistic samples of precipitation for wet days and to answer questions such as "What is the probability of getting more than 40 mm of precipitation today?," relevant for many impact activities. Key Points: We thoroughly assess different configuration options of the random forest technique for statistical downscaling of precipitationWe introduce a posteriori random forests, which reliably estimate the probability distribution of rainfall on wet daysAs a consequence, this new methodology allows us to generate realistic stochastic precipitation amounts without loss of predictive power [ABSTRACT FROM AUTHOR]

Statistical downscaling (SD) and bias adjustment (BA) methods are routinely used to produce regional to local climate change projections from coarse global model outputs. The suitability of these techniques depends on the particular application of interest and, especially, on the required spatial resolution. Whereas SD is appropriate for local (e.g., gauge) resolution, BA may be a good alternative when the gap between the predictor and predictand resolution is small. However, the different sources of uncertainty affecting SD such as reanalysis uncertainty, the choice of suitable predictors, climate model, and/or statistical approach may yield implausible projections in particular situations for which BA techniques may offer a compromise alternative, even for local resolution. In this work, we consider a case study with 41 rain gauges over Malawi and show that, despite producing similar results for a historical period, the use of different predictors may lead to large differences in the future projections obtained from SD methods. For instance, using temperature T (specific humidity Q) produces much drier (wetter) conditions than those projected by the raw global models for the target area. We demonstrate that this can be partially alleviated by substituting T+Q by relative humidity R, which simultaneously accounts for both water availability and temperature, and yields regional projections more compatible with the global one. Nevertheless, large local differences still persist, lacking a physical interpretation. In these situations, the use of simpler approaches such as empirical BA may lead to more plausible (i.e., more consistent with the global model) projections. [ABSTRACT FROM AUTHOR]

VALUE is an open European collaboration to intercompare downscaling approaches for climate change research, focusing on different validation aspects (marginal, temporal, extremes, spatial, process‐based, etc.). Here we describe the participating methods and first results from the first experiment, using "perfect" reanalysis (and reanalysis‐driven regional climate model (RCM)) predictors to assess the intrinsic performance of the methods for downscaling precipitation and temperatures over a set of 86 stations representative of the main climatic regions in Europe. This study constitutes the largest and most comprehensive to date intercomparison of statistical downscaling methods, covering the three common downscaling approaches (perfect prognosis, model output statistics—including bias correction—and weather generators) with a total of over 50 downscaling methods representative of the most common techniques. Overall, most of the downscaling methods greatly improve (reanalysis or RCM) raw model biases and no approach or technique seems to be superior in general, because there is a large method‐to‐method variability. The main factors most influencing the results are the seasonal calibration of the methods (e.g., using a moving window) and their stochastic nature. The particular predictors used also play an important role in cases where the comparison was possible, both for the validation results and for the strength of the predictor–predictand link, indicating the local variability explained. However, the present study cannot give a conclusive assessment of the skill of the methods to simulate regional future climates, and further experiments will be soon performed in the framework of the EURO‐CORDEX initiative (where VALUE activities have merged and follow on). Finally, research transparency and reproducibility has been a major concern and substantive steps have been taken. In particular, the necessary data to run the experiments are provided at http://www.value‐cost.eu/data and data and validation results are available from the VALUE validation portal for further investigation: http://www.value‐cost.eu/validationportal. [ABSTRACT FROM AUTHOR]

Statistical downscaling methods (SDMs) are techniques used to downscale and/or bias‐correct climate model results to regional or local scales. The European network VALUE developed a framework to evaluate and inter‐compare SDMs. One of VALUE's experiments is the perfect predictor experiment that uses reanalysis predictors to isolate downscaling skill. Most evaluation papers for SDMs employ simple statistical diagnostics and do not follow a process‐based rationale. Thus, in this paper, a process‐based evaluation has been conducted for the more than 40 participating model output statistics (MOS, mostly bias correction) and perfect prognosis (PP) methods, for temperature and precipitation at 86 weather stations across Europe. The SDMs are analysed following the so‐called "regime‐oriented" technique, focussing on relevant features of the atmospheric circulation at large to local scales. These features comprise the North Atlantic Oscillation, blocking and selected Lamb weather types and at local scales the bora wind and the western Iberian coastal‐low level jet. The representation of the local weather response to the selected features depends strongly on the method class. As expected, MOS is unable to generate process sensitivity when it is not simulated by the predictors (ERA‐Interim). Moreover, MOS often suffers from an inflation effect when a predictor is used for more than one station. The PP performance is very diverse and depends strongly on the implementation. Although conditioned on predictors that typically describe the large‐scale circulation, PP often fails in capturing the process sensitivity correctly. Stochastic generalized linear models supported by well‐chosen predictors show improved skill to represent the sensitivities. [ABSTRACT FROM AUTHOR]

In a previous study Pons et al. (Clim Res 54(3):197-207, . doi:) reported a significant decreasing trend of snowfall occurrence in the Northern Iberian Peninsula since the mid 70s. The study was based on observations of annual snowfall frequency (measured as the annual number of snowfall days NSD) from a network of 33 stations ranging from 60 to 1350 m. In the present work we analyze the skill of Regional Climate Models (RCMs) to reproduce this trend for the period 1961-2000 (using both reanalysis- and historical GCM-driven boundary conditions) and the trend and the associated uncertainty of the regional future projections obtained under the A1B scenario for the first half of the twenty-first century. In particular, we consider the regional simulation dataset from the EU-funded ENSEMBLES project, consisting of thirteen state-of-the-art RCMs run at 25 km resolution over Europe. While ERA40 severely underestimates both the mean NSD and its observed trend (−2.2 days/decade), the corresponding RCM simulations driven by the reanalysis appropriately capture the interannual variability and trends of the observed NSD (trends ranging from −3.4 to −0.7, −2.1 days/decade for the ensemble mean). The results driven by the GCM historical runs are quite variable, with trends ranging from −8.5 to 0.2 days/decade (−1.5 days/decade for the ensemble mean), and the greatest uncertainty by far being associated with the particular GCM used. Finally, the trends for the future 2011-2050 A1B runs are more consistent and significant, ranging in this case from −3.7 to −0.5 days/decade (−2.0 days/decade for the ensemble mean), indicating a future significant decreasing trend. These trends are mainly determined by the increasing temperatures, as indicated by the interannual correlation between temperature and NSD (−0.63 in the observations), which is preserved in both ERA40- and GCM-driven simulations. [ABSTRACT FROM AUTHOR]

Limited area models (LAMs) are widely used tools to downscale the wind speed forecasts issued by general circulation models. However, only a few studies have systematically analyzed the value added by the LAMs to the coarser-resolution-model wind. The goal of the present work is to investigate how added value depends on the resolution of the driving global model. With this aim, the Weather Research and Forecasting (WRF) Model was used to downscale three different global datasets (GFS, ERA-Interim, and NCEP-NCAR) to a 9-km-resolution grid for a 1-yr period. Model results were compared with a large set of surface observations, including land station and offshore buoy data. Substantial biases were found at this resolution over mountainous terrain, and a slight modification to the subgrid orographic drag parameterization was introduced to alleviate the problem. It was found that, at this resolution, WRF is able to produce significant added value with respect to the NCEP-NCAR reanalysis and ERA-Interim but only a small amount of added value with respect to GFS forecasts. Results suggest that, as model resolution increases, traditional skill scores tend to saturate. Thus, adding value to high-resolution global models becomes significantly more difficult. [ABSTRACT FROM AUTHOR]

This work shows that local-scale climate projections obtained by means of statistical downscaling are sensitive to the choice of reanalysis used for calibration. To this aim, a generalized linear model (GLM) approach is applied to downscale daily precipitation in the Philippines. First, the GLMs are trained and tested separately with two distinct reanalyses (ERA-Interim and JRA-25) using a cross-validation scheme over the period 1981-2000. When the observed and downscaled time series are compared, the attained performance is found to be sensitive to the reanalysis considered if climate change signal-bearing variables (temperature and/or specific humidity) are included in the predictor field. Moreover, performance differences are shown to be in correspondence with the disagreement found between the raw predictors from the two reanalyses. Second, the regression coefficients calibrated either with ERA-Interim or JRA-25 are subsequently applied to the output of a global climate model (MPI-ECHAM5) in order to assess the sensitivity of local-scale climate change projections (up to 2100) to reanalysis choice. In this case, the differences detected in present climate conditions are considerably amplified, leading to 'delta-change' estimates differing by up to 35% (on average for the entire country) depending on the reanalysis used for calibration. Therefore, reanalysis choice is an important contributor to the uncertainty of local-scale climate change projections and, consequently, should be treated with as much care as other better-known sources of uncertainty (e.g., the choice of the GCM and/or downscaling method). Implications of the results for the entire tropics, as well as for the model output statistics downscaling approach are also briefly discussed. [ABSTRACT FROM AUTHOR]

Climate Impact Indices (CIIs) are being increasingly used in different socioeconomic sectors to transfer information about climate change impacts to stakeholders. Typically, CIIs comprise into a single index several weather variables -such as temperature, wind speed, precipitation and humidity- which are relevant for a particular problem of interest. Moreover, most of the CIIs require daily (or monthly) physical coherence among these variables for their proper calculation. This constraints the number of statistical downscaling techniques suitable for a component-wise approach to this problem. We test the suitability of the alternative 'direct' downscaling approach in which the downscaling method is applied directly to the CII, thus circumventing the multi-variable problem and allowing the use of a wider range of downscaling methods. For illustrative purposes, we consider two popular CIIs -the Fire Weather Index (FWI) and the Physiological Equivalent Temperature (PET), used in the wildfire and tourism sectors, respectively- and compare the performance of the two approaches using the analog method, a simple and popular method providing inter-variable dependence. The results obtained with 'perfect' reanalysis predictors are comparable for both approaches, although smaller accuracy is obtained in general with the direct approach. Moreover, similar climate change 'deltas' are obtained with both approaches when applied to an illustrative future global projection using the ECHAM5 model. Overall, there is a trade-off between performance and simplicity which needs to be balanced for each particular application. [ABSTRACT FROM AUTHOR]

The effect of climate change on wildfires constitutes a serious concern in fire-prone regions with complex fire behavior such as the Mediterranean. The coarse resolution of future climate projections produced by General Circulation Models (GCMs) prevents their direct use in local climate change studies. Statistical downscaling techniques bridge this gap using empirical models that link the synoptic-scale variables from GCMs to the local variables of interest (using e.g. data from meteorological stations). In this paper, we investigate the application of statistical downscaling methods in the context of wildfire research, focusing in the Canadian Fire Weather Index (FWI), one of the most popular fire danger indices. We target on the Iberian Peninsula and Greece and use historical observations of the FWI meteorological drivers (temperature, humidity, wind and precipitation) in several local stations. In particular, we analyze the performance of the analog method, which is a convenient first choice for this problem since it guarantees physical and spatial consistency of the downscaled variables, regardless of their different statistical properties. First we validate the method in perfect model conditions using ERA-Interim reanalysis data. Overall, not all variables are downscaled with the same accuracy, with the poorest results (with spatially averaged daily correlations below 0.5) obtained for wind, followed by precipitation. Consequently, those FWI components mostly relying on those parameters exhibit the poorest results. However, those deficiencies are compensated in the resulting FWI values due to the overall high performance of temperature and relative humidity. Then, we check the suitability of the method to downscale control projections (20C3M scenario) from a single GCM (the ECHAM5 model) and compute the downscaled future fire danger projections for the transient A1B scenario. In order to detect problems due to non-stationarities related to climate change, we compare the results with those obtained with a Regional Climate Model (RCM) driven by the same GCM. Although both statistical and dynamical projections exhibit a similar pattern of risk increment in the first half of the 21st century, they diverge during the second half of the century. As a conclusion, we advocate caution in the use of projections for this last period, regardless of the regionalization technique applied. [ABSTRACT FROM AUTHOR]

The present study assesses the ability of seven Earth System Models (ESMs) from the Coupled Model Intercomparison Project Phase 5 to reproduce present climate conditions in Europe and Africa. This is done from a downscaling perspective, taking into account the requirements of both statistical and dynamical approaches. ECMWF's ERA-Interim reanalysis is used as reference for an evaluation of circulation, temperature and humidity variables on daily timescale, which is based on distributional similarity scores. To additionally obtain an estimate of reanalysis uncertainty, ERA-Interim's deviation from the Japanese Meteorological Agency JRA-25 reanalysis is calculated. Areas with considerable differences between both reanalyses do not allow for a proper assessment, since ESM performance is sensitive to the choice of reanalysis. For use in statistical downscaling studies, ESM performance is computed on the grid-box scale and mapped over a large spatial domain covering Europe and Africa, additionally highlighting those regions where significant distributional differences remain even for the centered/zero-mean time series. For use in dynamical downscaling studies, performance is specifically assessed along the lateral boundaries of the three CORDEX domains defined for Europe, the Mediterranean Basin and Africa. [ABSTRACT FROM AUTHOR]

The performance of statistical downscaling (SD) techniques is critically reassessed with respect to their robust applicability in climate change studies. To this end, in addition to standard accuracy measures and distributional similarity scores, the authors estimate the robustness of the methods under warming climate conditions working with anomalous warm historical periods. This validation framework is applied to intercompare the performances of 12 different SD methods (from the analog, weather typing, and regression families) for downscaling minimum and maximum temperatures in Spain. First, a calibration of these methods is performed in terms of both geographical domains and predictor sets; the results are highly dependent on the latter, with optimum predictor sets including near-surface temperature data (in particular 2-m temperature), which appropriately discriminate cold episodes related to temperature inversion in the lower troposphere. Although regression methods perform best in terms of correlation, analog and weather generator approaches are more appropriate for reproducing the observed distributions, especially in case of wintertime minimum temperature. However, the latter two families significantly underestimate the temperature anomalies of the warm periods considered in this work. This underestimation is found to be critical when considering the warming signal in the late twenty-first century as given by a global climate model [the ECHAM5-Max Planck Institute (MPI) model]. In this case, the different downscaling methods provide warming values with differences in the range of 1°C, in agreement with the robustness significance values. Therefore, the proposed test is a promising technique for detecting lack of robustness in statistical downscaling methods applied in climate change studies. [ABSTRACT FROM AUTHOR]

In this study, a worldwide overview on the expected sensitivity of downscaling studies to reanalysis choice is provided. To this end, the similarity of middle-tropospheric variables-which are important for the development of both dynamical and statistical downscaling schemes-from 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40) and NCEP-NCAR reanalysis data on a daily time scale is assessed. For estimating the distributional similarity, two comparable scores are used: the two-sample Kolmogorov-Smirnov statistic and the probability density function (PDF) score. In addition, the similarity of the day-to-day sequences is evaluated with the Pearson correlation coefficient. As the most important results demonstrated, the PDF score is found to be inappropriate if the underlying data follow a mixed distribution. By providing global similarity maps for each variable under study, regions where reanalysis data should not assumed to be 'perfect' are detected. In contrast to the geopotential and temperature, significant distributional dissimilarities for specific humidity are found in almost every region of the world. Moreover, for the latter these differences not only occur in the mean, but also in higher-order moments. However, when considering standardized anomalies, distributional and serial dissimilarities are negligible over most extratropical land areas. Since transformed reanalysis data are not appropriate for regional climate models-in opposition to statistical approaches-their results are expected to be more sensitive to reanalysis choice. [ABSTRACT FROM AUTHOR]