ALBERT

Description: Deep learning techniques (in particular convolutional neural networks, CNNs) have recently emerged as a promising approach for statistical downscaling due to their ability to learn spatial features from huge spatiotemporal datasets. However, existing studies are based on complex models, applied to particular case studies and using simple validation frameworks, which makes a proper assessment of the (possible) added value offered by these techniques difficult. As a result, these models are usually seen as black boxes, generating distrust among the climate community, particularly in climate change applications. In this paper we undertake a comprehensive assessment of deep learning techniques for continental-scale statistical downscaling, building on the VALUE validation framework. In particular, different CNN models of increasing complexity are applied to downscale temperature and precipitation over Europe, comparing them with a few standard benchmark methods from VALUE (linear and generalized linear models) which have been traditionally used for this purpose. Besides analyzing the adequacy of different components and topologies, we also focus on their extrapolation capability, a critical point for their potential application in climate change studies. To do this, we use a warm test period as a surrogate for possible future climate conditions. Our results show that, while the added value of CNNs is mostly limited to the reproduction of extremes for temperature, these techniques do outperform the classic ones in the case of precipitation for most aspects considered. This overall good performance, together with the fact that they can be suitably applied to large regions (e.g., continents) without worrying about the spatial features being considered as predictors, can foster the use of statistical approaches in international initiatives such as Coordinated Regional Climate Downscaling Experiment (CORDEX).

Description: The increasing demand for high-resolution climate information has attracted growing attention to statistical downscaling (SDS) methods, due in part to their relative advantages and merits as compared to dynamical approaches (based on regional climate model simulations), such as their much lower computational cost and their fitness for purpose for many local-scale applications. As a result, a plethora of SDS methods is nowadays available to climate scientists, which has motivated recent efforts for their comprehensive evaluation, like the VALUE initiative (http://www.value-cost.eu, last access: 29 March 2020). The systematic intercomparison of a large number of SDS techniques undertaken in VALUE, many of them independently developed by different authors and modeling centers in a variety of languages/environments, has shown a compelling need for new tools allowing for their application within an integrated framework. In this regard, downscaleR is an R package for statistical downscaling of climate information which covers the most popular approaches (model output statistics – including the so-called “bias correction” methods – and perfect prognosis) and state-of-the-art techniques. It has been conceived to work primarily with daily data and can be used in the framework of both seasonal forecasting and climate change studies. Its full integration within the climate4R framework (Iturbide et al., 2019) makes possible the development of end-to-end downscaling applications, from data retrieval to model building, validation, and prediction, bringing to climate scientists and practitioners a unique comprehensive framework for SDS model development. In this article the main features of downscaleR are showcased through the replication of some of the results obtained in VALUE, placing an emphasis on the most technically complex stages of perfect-prognosis model calibration (predictor screening, cross-validation, and model selection) that are accomplished through simple commands allowing for extremely flexible model tuning, tailored to the needs of users requiring an easy interface for different levels of experimental complexity. As part of the open-source climate4R framework, downscaleR is freely available and the necessary data and R scripts to fully replicate the experiments included in this paper are also provided as a companion notebook.

Description: The present work presents a new observational gridded dataset (referred to as Iberia01) for daily precipitation and temperatures produced using a dense network (thousands) of stations over the Iberian Peninsula for the period 1971–2015 at 0.1∘ regular (and 0.11∘ CORDEX-compliant rotated) resolutions. We analyze mean and extreme indices and compare the results with the E-OBS v17 dataset (using both the standard and ensemble versions, at 0.25 and 0.1∘ resolutions, respectively), in order to assess observational uncertainty in this region. We show that Iberia01 produces more realistic precipitation patterns than E-OBS for the mean and extreme indices considered, although both are comparable for temperatures. To assess the differences between these datasets, a new probabilistic intercomparison analysis was performed, using the E-OBS ensemble (v17e) to characterize observational uncertainty and testing whether Iberia01 falls within the observational uncertainty range provided by E-OBS. In general, uncertainty values are large in the whole territory, with the exception of a number of kernels where the uncertainty is small, corresponding to the stations used to build the E-OBS grid. For precipitation, significant differences – at the 10 % level – between both datasets were found for fewer than 25 % of days over the Iberian Peninsula. For temperature, a very inhomogeneous spatial pattern was obtained, with either a small (in most of the regions) or large fraction of significantly different days, thus indicating sensible regions for observational uncertainty. Iberia01 is publicly available (Herrera et al., 2019a, https://doi.org/10.20350/digitalCSIC/8641).

Description: The increasing demand for high-resolution climate information has attracted a growing attention for statistical downscaling methods (SD), due in part to their relative advantages and merits as compared to dynamical approaches (based on regional climate model simulations), such as their much lower computational cost and their fitness-for-purpose for many local-scale applications. As a result, a plethora of SD methods is nowadays available for climate scientists, which has motivated recent efforts for their comprehensive evaluation, like the VALUE Project (http://www.value-cost.eu). The systematic intercomparison of a large number of SD techniques undertaken in VALUE, many of them independently developed by different authors and modeling centers in a variety of languages/environments, has shown a compelling need for new tools allowing for their application within an integrated framework. With this regard, downscaleR is an R package for statistical downscaling of climate information which covers the most popular approaches (Model Output Statistics – including the so called 'bias correction' methods – and Perfect Prognosis) and state-of-the-art techniques. It has been conceived to work primarily with daily data and can be used in the framework of both seasonal forecasting and climate change studies. Its full integration within the climate4R framework (Iturbide et al. 2019) makes possible the development of end-to-end downscaling applications, from data retrieval to model building, validation and prediction, bringing to climate scientists and practitioners a unique comprehensive framework for SD model development. In this article the main features of downscaleR are showcased through the replication of some of the results obtained in the VALUE Project, making an emphasis in the most technically complex stages of perfect-prog model calibration (predictor screening, cross-validation and model selection) that are accomplished through simple commands allowing for extremely flexible model tuning, tailored to the needs of users requiring an easy interface for different levels of experimental complexity. As part of the open-source climate4R framework, downscaleR is freely available and the necessary data and R scripts to fully replicate the experiments included in this paper are also provided as a companion notebook.

Description: The present work introduces a new observational gridded dataset produced using a dense network (thousands) of stations over the Iberian Peninsula (referred to as Iberia01, Gutiérrez et al. (2019), DOI: http://dx.doi.org/10.20350/digitalCSIC/8641), providing daily precipitation and temperatures for the period 1971–2015 at 0.1° regular (and 0.11° rotated CORDEX compliant) resolutions. A comparison with both the standard and ensemble version of the E-OBS v17 dataset (at 0.25° and 0.1° resolutions, respectively) is undertaken in order to assess observational uncertainty in this region. First, a standard comparison is performed for several weather indices, obtaining the differences between both datasets. Secondly, a new probabilistic intercomparison analysis is introduced, using the E-OBS ensemble (v17e) to characterize observational uncertainty and testing the hypothesis that Iberia01 is a realization of the ensemble (i.e. it falls within the observational uncertainty range provided by E-OBS). Finally, the effective resolution of the auxiliar very high resolution grid (0.01°) built to obtain the area-average representativity of the final dataset, and thus the possibility to increase the resolution of the dataset by means of pure interpolation methods, is analyzed considering an extreme event of convective precipitation affecting the Iberian Peninsula. As a result, we show that Iberia01 produces more realistic patterns than E-OBS v17 in the case of precipitation for all the indices considered, although both are comparable for temperatures. These differences are assessed using the probabilistic approach based on the E-OBS ensemble showing a quite homogeneous spatial pattern for precipitation (with less than 25 % significantly – at a 10 % level – different days between both datasets) and a very inhomogeneous pattern for temperatures, with either a small (in most of the regions) or large fraction of significantly different days. The great uncertainty of the precipitation given by E-OBS ensemble, in which the standard deviation of the ensemble has the same order than the mean value, increases the significance of the results obtained for this variable reflecting the differences between both datasets.

Description: Several sets of reference regions have been used in the literature for the regional synthesis of observed and modelled climate and climate change information. A popular example is the series of reference regions used in the Intergovernmental Panel on Climate Change (IPCC) Special Report on Managing the Risks of Extreme Events and Disasters to Advance Climate Adaptation (SREX). The SREX regions were slightly modified for the Fifth Assessment Report of the IPCC and used for reporting subcontinental observed and projected changes over a reduced number (33) of climatologically consistent regions encompassing a representative number of grid boxes. These regions are intended to allow analysis of atmospheric data over broad land or ocean regions and have been used as the basis for several popular spatially aggregated datasets, such as the Seasonal Mean Temperature and Precipitation in IPCC Regions for CMIP5 dataset. We present an updated version of the reference regions for the analysis of new observed and simulated datasets (including CMIP6) which offer an opportunity for refinement due to the higher atmospheric model resolution. As a result, the number of land and ocean regions is increased to 46 and 15, respectively, better representing consistent regional climate features. The paper describes the rationale for the definition of the new regions and analyses their homogeneity. The regions are defined as polygons and are provided as coordinates and a shapefile together with companion R and Python notebooks to illustrate their use in practical problems (e.g. calculating regional averages). We also describe the generation of a new dataset with monthly temperature and precipitation, spatially aggregated in the new regions, currently for CMIP5 and CMIP6, to be extended to other datasets in the future (including observations). The use of these reference regions, dataset and code is illustrated through a worked example using scatter plots to offer guidance on the likely range of future climate change at the scale of the reference regions. The regions, datasets and code (R and Python notebooks) are freely available at the ATLAS GitHub repository: https://github.com/SantanderMetGroup/ATLAS (last access: 24 August 2020), https://doi.org/10.5281/zenodo.3998463 (Iturbide et al., 2020).