Abstract

This study compares three machine learning (ML) models [long short-term memory (LSTM), random forest (RF), and XGBoost] and four satellite altimetry missions (J3, S3B, S6, and SWOT) for estimating water levels in Lake Gray, Chilean Patagonia (2012-2025). XGBoost achieved the best ML performance (RMSE = 0.11 +/- 0.02 m, Kling-Gupta efficiency (KGE) = 0.9684 +/- 0.0084), while Sentinel-6 provided the highest altimetric accuracy (RMSE = 0.07 m). SWOT captured the full hydrological variability (IQR = 0.78 m), which was nearly identical to in situ records, outperforming traditional altimeters (J3 and S3B). Tree-based methods and SWOT best reproduce observed distributions. The combined use of ML for longterm reconstruction and modern altimetry for high-precision monitoring offers a reproducible, proof-of-concept framework for remote, ice-influenced proglacial lakes.