Applications of Deep Learning to S2S Precipitation Prediction and Downscaling for the Middle East and North Africa

Machine learning algorithms are changing the world, and the study of hydroclimatic variability and predictability is no exception. Precipitation prediction and downscaling have emerged as an area of potential for the application of Convolutional Neural Networks (CNN), which are deep learning algorithms that offer several advantages. CNN can automatically extract features from input data, capture nonlinear responses and interactions, and be optimized with a custom “regional” loss function, using the results of objective regionalization, allowing for different predictive structures in different regions. In this application, we implement CNN using observed data and dynamical forecasts at subseasonal to seasonal (S2S) timescales for different countries in the Middle East and North Africa (MENA). Training is performed using predictors from historical observations and the North American Multi-Model Ensemble (NMME) forecasts along with gridded 5-km resolution CHIRPS precipitation as predictand. A custom loss function is defined to minimize the root mean squared error (RMSE) of the regional average for each defined climate region. We use CNN to generate high-resolution gridded precipitation predictions at monthly time scale. The proposed CNN workflow has been successfully applied to improve the skill and resolution of dynamically-based S2S precipitation forecasts for MENA.