• English 
  • Spanish 

Deep Convolutional Networks for Feature Selection in Statistical Downscaling

The potential of (deep) convolutional neural networks for automatic predictor selection in statistical downscaling over large continental domains is analyzed focusing on a simple illustrative example (precipitation occurrence). It is shown that these models automatically handle redundancy and perform geographical and variable selection/transformation of predictors in a robust and spatially consistent form, obtaining similar features for different predictor sets. Results are compared with best performing standard methods from the largest-to-date intercomparison of statistical downscaling methods (VALUE) using “perfect” reanalysis predictors.