Advancing High‐Resolution Lake Bathymetry Reconstruction Through Geomorphologically Informed Deep Learning in High Mountain Asia

Abstract

Accurate three-dimensional (3D) lake bathymetry reconstruction is critical for water resources assessment and hydrological modeling yet remains constrained by data scarcity and oversimplified geometric assumptions. To address these challenges, we propose the Geomorphologically informed deep learning (GIDL) framework for high-resolution 3D lake bathymetry reconstruction. GIDL resolves the critical data bottleneck by constructing abundant virtual lake samples from terrestrial river valleys and depressions, which serve as geomorphological analogs to underwater terrain. The framework integrates geomorphology-based mathematical priors, which encode terrain continuity, slope-related depth variation, and basin-shape regularities, to establish a preliminary basin skeleton. Trained on the geomorphologically consistent virtual samples, the network learns to identify and correct the systematic deviations of the mathematical approximation, thereby recovering high-fidelity morphological details. To further refine performance, lakes are classified using a Rectanglarity and Ellipticity Index, enabling morphology-specific learning. Validated in High Mountain Asia (HMA) using 18,753 training and 4,688 independent virtual samples for validation and testing, GIDL achieved reliable bathymetric reconstructions (mean NRMSE: 16.14%, R
²: 0.48), reducing errors by ∼35% compared to mathematical priors, and outperforming deep learning baselines. Extensive testing on 30 real HMA lakes demonstrated more than a 60% reduction in depth estimation errors compared to existing state-of-the-art models. Furthermore, application to 605 lakes yielded robust Area-Depth-Volume relationships (R
² = 0.75). By explicitly incorporating geomorphic knowledge into data-driven modeling, GIDL offers a transferable methodological paradigm for reconstructing ungauged lake terrains, significantly advancing global water resource management and climate monitoring.

Source link