The Spanish Meteorological Agency (AEMET) is responsible for the elaboration of downscaled climate projections over Spain to feed the Second National Plan of Adaptation to Climate Change (PNACC-2). The main objective of this article is to establish a comparison among five statistical downscaling methods developed at AEMET: (1) Analog, (2) Regression, (3) Artificial Neural Networks, (4) Support Vector Machines and (5) Kernel Ridge Regression. This comparison has been carried out under present conditions and with perfect predictors, based on the framework established by the VALUE network, in particular, on its perfect predictor experiment. In this experiment, we evaluate the marginal aspects of the distributions of daily maximum/minimum temperatures and daily accumulated precipitation analysed by seasons, on a high resolution observational grid (0.05 ) over mainland Spain and the Balearic Islands. This is the first of a set of three experiments aimed to allow us to decide which methods, and under what configuration, is more appropriate for the generation of downscaled climate projections over our region. For maximum/minimum temperatures, all methods display a similar behaviour. They capture very satisfactorily the mean values although slight biases are detected on the extremes. In general, results for maximum temperature appear to be more accurate than for minimum temperature, and the nonlinear methods display certain added value. For precipitation, remarkable differences are found among all methods. Most of the methods are capable of reproducing the total precipitation amount quite satisfactorily, whereas other aspects such as intense precipitations and the precipitation occurrence are captured with more accuracy by the Analog method.

Machine learning is a growing field of research with many applications. It provides a series of techniques able to solve complex nonlinear problems, and that has promoted their application for statistical downscaling. Intercomparison exercises with other classical methods have so far shown promising results. Nevertheless, many evaluation studies of statistical downscaling methods neglect the analysis of their extrapolation capability. In this study, we aim to make a wakeup call to the community about the potential risks of using machine learning for statistical downscaling of climate change projections. We

Abstract. The Pyrenees, located in the transition zone of Atlantic and Mediterranean climates, constitute a paradigmatic example of mountains undergoing rapid changes in environmental conditions, with potential impact on the availability of water resources, mainly for downstream populations. High-resolution probabilistic climate change projections for precipitation and temperature are a crucial element for stakeholders to make well-informed decisions on adaptation to new climate conditions. In this line, we have generated high–resolution climate projections for 21st century by applying two statistical downscaling methods (regression for max and min temperatures, and analogue for precipitation) over the Pyrenees region in the frame of the CLIMPY project over a new high-resolution (5 km × 5 km) observational grid using 24 climate models from CMIP5. The application of statistical downscaling to such a high resolution observational grid instead of station data partially circumvent the problems associated to the non-uniform distribution of observational in situ data. This new high resolution projections database based on statistical algorithms complements the widely used EUROCORDEX data based on dynamical downscaling and allows to identify features that are dependent on the particular downscaling method. In our analysis, we not only focus on maximum and minimum temperatures and precipitation changes but also on changes in some relevant extreme indexes, being 1986–2005 the reference period. Although climate models predict a general increase in temperature extremes for the end of the 21st century, the exact spatial distribution of changes in temperature and much more in precipitation remains uncertain as they are strongly model dependent. Besides, for precipitation, the uncertainty associated to models can mask – depending on the zones- the signal of change. However, the large number of downscaled models and the high resolution of the used grid allow us to provide differential information at least at massif level. The impact of the RCP becomes significant for the second half of the 21st century, with changes – differentiated by massifs – of extreme temperatures and analysed associated extreme indexes for RCP8.5 at the end of the century.

<p>The Spanish Meteorological Agency (AEMET) is responsible for the elaboration of downscaled climate projections over Spain to feed the Second National Plan of Adaptation to Climate Change (PNACC-2). The main objective of this work is to establish a comparison among five statistical downscaling methods developed at AEMET: 1) Analog, 2) Regression, 3) Artificial Neural Networks, 4) Support Vector Machines and 5) Kernel Ridge Regression. All the five methods have been applied with a Perfect Prog approach to downscale daily maximum/minimum temperatures and daily precipitation on a high resolution observational grid (0.05<sup>o</sup>) over mainland Spain and the Balearic Islands, a region particularly challenging due to its large regional spatio-temporal variabilities. The comparison has been carried out under present conditions and with perfect predictors, based on the framework established by the VALUE network, in particular, on its perfect predictor experiment. The evaluation here performed is focused on marginal aspects, through an analysis of the four seasonal distributions of each variable. In order to enable a manageable comparison among all methods three indexes commonly used for climate change adaptation and impact studies have been used: the mean value and the 10th and 90th percentiles for daily maximum/minimum temperatures and the total precipitation amount (PRCPTOT), the total precipitation on very wet days (R95p) and the number of wet days (R01) for precipitation. For maximum/minimum temperatures, all methods display a similar behavior. They capture very satisfactorily the mean values although slight biases are detected on the extremes. In general, results for maximum temperature appear to be more accurate than for minimum temperature, and the non-linear methods display certain added value. For precipitation, remarkable differences are found among all methods: most of them are capable of reproducing the total precipitation amount quite satisfactorily, while other aspects such as intense precipitations and the precipitation occurrence are captured with more accuracy by the Analog method.</p><p> </p>

The Spanish Meteorological Agency (AEMET) is responsible for the elaboration of downscaled climate projections over Spain to feed the Second National Plan of Adaptation to Climate Change (PNACC-2) and this is the last of three papers aimed to evaluate and intercompare five empirical/statistical downscaling (ESD) methods developed at AEMET: (a) Analog, (b) Regression, (c) Artificial Neural Networks, (d) Support Vector Machines and (e) Kernel Ridge Regression, in order to decide which methods and under what configurations are more suitable for that purpose. Following the framework established by the EU COST Action VALUE, in this experiment we test the transferability of these methods to future climate conditions with the use of regional climate models (RCMs) as pseudo observations. We evaluate the marginal aspects of the distributions of daily maximum/minimum temperatures and daily accumulated precipitation, over mainland Spain and the Balearic Islands, analysed by season. For maximum/minimum temperatures all methods display certain transferability issues, being remarkable for Support Vector Machines and Kernel Ridge Regression. For precipitation all methods appear to suffer from transferability difficulties as well, although conclusions are not as clear as for temperature, probably due to the fact that precipitation does not present such a marked signal of change. This study has revealed how an analysis over a historical period is not enough to fully evaluate ESD methods, so we propose that some type of analysis of transferability should be added in a standard procedure of a complete evaluation.