ARPEGE Cloud Cover Forecast Postprocessing with Convolutional Neural Network

Abstract Cloud cover provides crucial information for many applications such as planning land observation missions from space. It remains, however, a challenging variable to forecast, and numerical weather prediction (NWP) models suffer significant biases, hence, justifying the use of statistical postprocessing techniques. In this study, ARPEGE (Météo-France global NWP) cloud is postprocessed using convolutional neural network (CNN). CNN most popular machine learning tool deal with images. our case, allows integration spatial contained in NWP outputs. We gridded product derived satellite observations over Europe ground truth, predictors are fields various variables produced by at corresponding lead time. show that simple U-Net architecture (a particular type CNN) produces improvements Europe. Moreover, outclasses more traditional methods used operationally random forest logistic quantile regression. When large number predictors, first step toward interpretation produce ranking importance. Traditional (permutation importance, sequential selection, etc.) need important computational resources. introduced weighting predictor layer prior order ranking. The small additional weights train (the same predictors) does not impact time, representing huge advantage compared methods.