Explainable deep learning for insights in El Niño and river flows.

Area of Science:

• Climate Science
• Hydrology
• Machine Learning

Background:

• El Niño Southern Oscillation (ENSO) influences global hydrology via teleconnections.
• Deep Learning (DL) and Complex Networks (CN) show promise for ENSO prediction and teleconnection analysis.
• Gaps exist in understanding DL's black-box nature, simplified ENSO indices, and translating predictions to river flows.

Purpose of the Study:

• To develop eXplainable DL (XDL) methods combined with CN for improved ENSO-driven river flow prediction.
• To extract interpretable predictive information from global sea surface temperature (SST) data.
• To enhance the understanding of SST influences on river flows and provide uncertainty estimation.

Main Methods:

• Utilized eXplainable Deep Learning (XDL) with saliency maps to analyze global SST data.
• Integrated XDL with Complex Network (CN) constructions to understand teleconnections.
• Employed observations, reanalysis data, and earth system model simulations.

Main Results:

• Identified interpretable predictive information in global SST beyond standard ENSO indices.
• Discovered specific SST regions and dependence structures relevant to river flows.
• Achieved improved river flow predictions with uncertainty estimation.

Conclusions:

• XDL-CN methods offer enhanced predictive understanding of ENSO-driven river flows.
• The approach reveals additional information in SST and clarifies SST-river flow relationships.
• Demonstrated value for interannual and decadal scale climate projections.