Producing an operational and accurate seasonal forecast system is a complete game-changer for drylands regions grappling with severe climate variability and depleting freshwater resources. Having access to precise and accurate forecasts months in advance constitutes an extraordinary tool for policy-makers and farmers alike, and leads to better decision-making surrounding water allocation and conservation.

Seasonal forecasts are based on several model simulations: computer programs that are fed climate indices, observations on sea surface temperature, and El Niño/Southern Oscillation (ENSO), North Atlantic Oscillation (NAO), Mediterranean Oscillation (MO), and others.

One globally prominent seasonal forecast system is called the North American Multi-Model Ensemble (NMME). It consists of coupled models from several North American modeling centers. However, all NMME models show poor scores in predicting rainfall more than a month in advance when applied to the MENA region.

Our climate and water modeling team uses advanced techniques to improve the outputs of four NMME dynamic models (models (CFS, CanCM4, GEOS5, and GFDL-NEMO) in rainfall prediction. They fed these model outputs and CHIRPS precipitation data into a Convolutional Neural Network (CNN) to predict the rainfall at a two-month lead time. A CNN is an artificial neural network used in image recognition and processing that has high accuracy in processing pixel data. The loss function of the CNN model, re-written by the team, uses a “regionalization” technique, which consists in identifying coherent regions based on their specific rainfall regime. The team used the 1982-2013 period for training data and then tested and validated the deep learning CNN model outputs from the 2014-2019 period..

CNN model (top middle pane) captures the general distribution and intensity of higher rainfall events observed in the CHIRPS precipitation data (top left pane). It also shows how the CNN model output reflects the regional dry areas well. Overall, it is a significant improvement compared to the 4 NMME model outputs.
The modeling exercise was applied to Morocco, and it yielded quality results for predicting rainfall two months into the future.