8 September 2026 in SSRN Electronic Journal
Accurate supraglacial lake detection is essential for understanding ice sheet hydrology, yet effectively integrating topographic constraints into deep learning frameworks remains challenging. Existing fusion strategies treat optical and topographic data as homogeneous inputs and fail to enforce the physical constraints that terrain imposes on lake formation. This study proposes the Terrain-Guided Fusion Network (TGFSD), which treats ArcticDEM-derived features as a conditional guide rather than an equal input. Two complementary modules are introduced: an Asymmetric Gated Fusion module that adaptively modulates optical responses based on local terrain suitability, and an Optical-Topographic Co-Attention module that supplies long-range topographic context when local cues are insufficient. Evaluated on Sentinel-2 imagery in northeastern Greenland, TGFSD achieves an F1-score of 0.9126, outperforming UNet, PSPNet, DeepLabV3+, SegFormer, SwinUNet, and their early-fusion variants. Ablation experiments confirm that both modules contribute positively, and the model exhibits strong robustness under thin cloud cover and water-filled crevasses. Topographic analysis further reveals that supraglacial lakes occupy distinct terrain niches characterized by low slope and moderate Topographic Wetness Index, supporting the physical rationale of the gating design. These results demonstrate an effective AI-driven pathway for cross-modal data fusion in cryospheric remote sensing.