Explainable deep learning for insights in El Niño and river flows

This paper investigates the use of explainable deep learning (XDL) methods to improve the predictive understanding of El Niño-Southern Oscillation (ENSO)-driven river flows. The authors address limitations of traditional methods, such as the black box nature of deep learning and the reliance on simplistic ENSO indices. By using saliency maps, they extract interpretable predictive information from global sea surface temperature (SST) data, identifying relevant SST regions and dependence structures for improved river flow prediction, including uncertainty estimation. The study uses observations, reanalysis data, and earth system model simulations to demonstrate the effectiveness of the XDL approach in conjunction with climate network constructions.

Yumin Liu, Kate Duffy, Jennifer G. Dy, Auroop R. Ganguly