SAE-Xplainers Interpret Extreme Earth Events

Researchers have developed a new interpretability framework called SAE-Xplainers to make deep learning models for extreme Earth events (ExEE) more understandable. While deep learning excels at processing large weather and climate datasets, its "black box" nature hinders adoption in operational settings. SAE-Xplainers adapt Sparse Autoencoders (SAEs), typically used for text and image, to the unique challenges of W&C data. The method involves two key innovations: first, a geographic location-based modulation of SAE inputs helps capture the local semantic meaning of environmental patterns. Second, an ensemble of rule-based SAE-Xplainers interprets the high-dimensional features derived from multi-modal environmental predictors. This approach has been validated on predicting fires and detecting tropical cyclones and atmospheric rivers, demonstrating improved reconstruction performance and the ability to generate human-understandable rules consistent with scientific literature.