Mapping Antarctic Blue Ice Areas With Sentinel-2A/B Images and LightGBM Model

Abstract

Antarctic blue ice plays a crucial role in surface energy balance and paleoclimate research. A high-accuracy and comprehensive dataset of blue ice areas (BIAs) is essential for understanding climate dynamics and environmental changes in the region. While satellite remote sensing is effective in mapping BIAs, traditional methods rely on limited spectral bands and linear models with inherent limitations. This study integrated remote sensing techniques with ensemble learning algorithms to develop a high-resolution (10 m) Antarctic-wide BIA dataset using Sentinel-2 imagery based on the years 2017–2022. Random forest, XGBoost, and LightGBM integrated learning algorithms were used to model the extraction of Antarctic blue ice. The accuracy of the model was evaluated by confusion matrix with LightGBM achieving the highest overall accuracy (87.23%). We also used SHapley Additive exPlanations values to improve the interpretability of opaque system models by evaluating the contribution of each feature variable. Validation through visual interpretation of Sentinel-2A/B images further confirmed the model's reliability, with an accuracy of 90.61%. Based on these robust results, we generated detailed BIAs across Antarctica. Our findings estimate the total BIAs at 175 274 km², covering approximately 1.25% of the continent. The blue ice is mainly concentrated in low-elevation coastal areas and mountain slopes, especially in Dronning Maud Land, Amery Ice Shelf, Wilkes Land, Victoria Land, and Transantarctic Mountains. We further reveal that most of the blue ice is located at elevations below 500 m, with air temperatures between −5 and 0 °C, and ice velocity under 100 m/yr. Our high-resolution dataset provides crucial insights for future research in Antarctic glaciology, paleoclimate studies, and meteorite collection.