In a recent paper, Baño-Medina et al. (Configuration and Intercomparison of deep learning neural models for statistical downscaling. preprint, 2019) assessed the suitability of deep convolutional neural networks (CNNs) for downscaling of temperature and precipitation over Europe using large-scale ‘perfect’ reanalysis predictors. They compared the results provided by CNNs with those obtained from a set of standard methods which have been traditionally used for downscaling purposes (linear and generalized linear models), concluding that CNNs are well suited for continental-wide applications. That analysis is extended here by assessing the suitability of CNNs for downscaling future climate change projections using Global Climate Model (GCM) outputs as predictors. This is particularly relevant for this type of “black-box” models, whose results cannot be easily explained based on physical reasons and could potentially lead to implausible downscaled projections due to uncontrolled extrapolation artifacts. Based on this premise, we analyze in this work the two key assumptions that are made in perfect prognosis downscaling: (1) the predictors chosen to build the statistical model should be well reproduced by GCMs and (2) the statistical model should be able to reliably extrapolate out of sample (climate change) conditions. As a first step to test the suitability of these models, the latter assumption is assessed here by analyzing how the CNNs affect the raw GCM climate change signal (defined as the difference, or delta, between future and historical climate). Our results show that, as compared to well-established generalized linear models (GLMs), CNNs yield smaller departures from the raw GCM outputs for the end of century, resulting in more plausible downscaling results for climate change applications. Moreover, as a consequence of the automatic treatment of spatial features, CNNs are also found to provide more spatially homogeneous downscaled patterns than GLMs.
