Speaker
Description
Antarctic Sea Ice (ASI) shows strong regional and interannual variability shaped by a complex interplay of atmospheric and oceanic processes. As the climate system redistributes energy poleward to offset radiative imbalances, remote conditions can influence ASI through large-scale circulation patterns, or teleconnections. Established teleconnection indices link these patterns to ASI variability via winds, heat fluxes, and moisture transport, but their spatial aggregation limits their ability to explain regional sea ice extremes — a limitation compounded by evidence that the dominant drivers differ from one extreme event to the next, underscoring the need for methods that can adapt to individual events rather than assuming a single fixed pattern governs ASI variability as a whole.
We propose an instance-wise feature attribution framework to address this gap. Rather than relying on predefined indices, we convert global climate fields into superpixels to preserve regional coherence, then train a selector network jointly with a predictor network so that each prediction is paired with a learned mask identifying the regions most responsible for it. This approach is designed to surface non-linear, spatially distributed, and event-specific drivers of ASI variability that static indices cannot capture.