Integrating ICESat-2 Lidar and Machine Learning for High-Precision Glacial Lake Volume Estimation in Southeastern Tibet Plateau

Glacial lakes formed during glacier retreat temporarily preserve meltwater and mitigate water resource pressure, dependent on ice reserves. However, this effect is transient, and regions relying on glacial meltwater may face future water scarcity. Due to the unique characteristics of maritime glaciers, glaciers in southeastern Tibet are experiencing the most significant mass loss on Earth—approximately three times the average rate across the Tibetan Plateau—exerting profound impacts on regional water resources. Understanding glacial lake evolution and mass balance in this region is therefore critical for developing early warning systems, mitigating glacial lake outburst flood risks, and safeguarding water resources.

Existing glacial lake studies primarily focus on monitoring area changes, while direct depth and volume observations remain extremely scarce due to challenging field conditions. Traditional empirical area-volume formulas inadequately capture morphological, topographic, and hydrological variations across lake types, resulting in significant estimation errors. Although ICESat-2 possesses shallow-water bathymetry capabilities, its application faces challenges, including limited single-beam coverage, significant basin information gaps, and constraints imposed by beam spacing, cloud cover, and variable turbidity.

This study systematically integrates ICESat-2 single-beam photon data with multi-source remote sensing and topographic-meteorological data, proposing a comprehensive framework progressing from "single laser profiles" to "three-dimensional basin reconstruction" and ultimately to "regional-scale glacial lake volume estimation." Based on ICESat-2 ATL03 geolocated photon data, we established a rigorous filtering workflow that combines quality flags, confidence constraints, and interactive manual selection to eliminate noise while retaining reliable bathymetric information. We developed a novel three-dimensional basin reconstruction model that optimizes bathymetry point distribution via mirror symmetry and contour-shrinkage mechanisms, with quadratic spline interpolation constraining the deepest point and radial basis functions enabling continuous terrain reconstruction.

Model validation using unmanned boat sonar measurements across multiple glacial lakes demonstrates that the proposed method stably reproduces bowl-shaped topographic features, with volume reconstruction errors generally below 10% and only 2% variation across different lake-bottom center assumptions, confirming robustness under complex observational conditions.

By integrating in-situ observations with ICESat-2 reconstructions, we constructed a high-quality dataset of 611 samples that incorporates lake morphology, topography, hydrology, and meteorology. Using Isolation Forest filtering and the XGBoost algorithm with optimized hyperparameters and recursive feature elimination, the model significantly outperforms traditional empirical formulas, achieving an R² of 0.911 for small and medium-small glacial lakes. SHAP analysis revealed lake area as the most critical variable, with lakeshore slope, shape regularity, and regional precipitation exerting significant regulatory effects. Monte Carlo uncertainty analysis demonstrates over 88% coverage of actual volumes within 95% confidence intervals with significantly lower bias than existing methods.

This study achieves a methodological breakthrough from ICESat-2 single-beam bathymetry to three-dimensional basin reconstruction, establishing a high-precision regional-scale estimation model. It provides a scalable technical framework for glacial lake hazard assessment, water resource monitoring, and the development of early warning systems in high-mountain regions.