Learning and Predicting Seasonal Glacier Dynamics from HELIX-Fused Sentinel-1 and TerraSAR-X Time Series

Abstract

Seasonal glacier facies evolve gradually, yet many SAR products still treat acquisitions as independent classification snapshots. This paper presents an interpretable and data-efficient framework for seasonal facies forecasting from open-access Sentinel-1 time series by enriching the supervision signal rather than increasing model complexity. TerraSAR-X facies maps are processed with HELIX, a label-side spatio-temporal enrichment pipeline that converts categorical facies into seasonally aggregated continuous targets capturing persistence and transitions. Temporally fused Sentinel-1 descriptors drive a multi-output gradient-boosted tree regressor predicting per-pixel seasonal facies tendencies, followed by residual correction and an optional historical prior. Experiments over Axel Heiberg Island and Devon Island across multiple years reduce MAE by >30% relative to a baseline and yield stable cross-year and cross-site transfer with sub-facies accuracy