An Integrated, Explainable and Uncertainty-Aware Satellite-Based Precipitation Product for the Mediterranean Region

Precipitation is a key variable of earth observation systems, involved in a wide spectrum of applications, from the optimal water resources management to early warning systems for flood/drought events. During the last decade, significant research effort has been given in the development of satellite-based precipitation products (SPPs), with improved accuracy and high spatio-temporal coverage. This work presents a methodological framework that couples machine learning and explainability techniques, along with uncertainty quantification methods, to develop a new SPP for the Mediterranean region. Specifically, the methodology couples: a) ensemble (tree-based) machine learning algorithms to merge multiple satellite-based observations (i.e., both precipitation and soil moisture-based products), b) global and local explainable AI techniques to provide explanations at the model and instance level, respectively, and c) a copula-based approach to quantify the predictive uncertainty of the final integrated product. The new product fuses data from two widely known low-latency SPPs, i.e., the SM2RAIN-ASCAT and GPM Late Run dataset, while its performance is assessed against state-of-art products, i.e., the EMO5, GPM Final Run, ERA5 and PERSIANN CDR.