Authors: Getnet Demil & Mourad Oussalah
Publication type: A1
Abstract
Accurate fractional snow cover (FSC) mapping requires combining the complementary strengths of optical imagery (spectral discrimination) and Synthetic Aperture Radar (all-weather capability). Yet existing fusion approaches employ simplistic concatenation strategies that fail to capture complex inter-modal relationships, while taking into account the resolution mismatch between high-resolution inputs and coarse reference labels. This paper introduces XFuse, a hybrid CNN-Transformer architecture that addresses these challenges through three key contributions: (1) bidirectional cross-attention fusion that explicitly models SAR-optical interdependencies via query-key–value transformations, enabling each modality to selectively attend to complementary information from the other; (2) a multi-resolution training strategy () that fuses Sentinel-1/2 features at native resolution, aggregates to 60 m for supervision against Gap-Filled Fractional Snow Cover (GFSC) labels, and disaggregates through guided upsampling to restore fine spatial structure at ; and (3) quality-aware curriculum learning that incorporates multi-source metadata (cloud masks, atmospheric parameters, quality control flags) through adaptive loss weighting, enabling robust training from heterogeneous satellite data while providing calibrated uncertainty estimates. Evaluation on 29 Sentinel-1/2 tiles from Northern Finland (18,117 test patches from 4 geographically unseen tiles) demonstrates state-of-the-art performance: 13.77% MAE and 94.66% classification accuracy, representing 38.7% MAE reduction over the best baseline. Ablation studies confirm cross-attention as the largest individual contributor (5.16 percentage point accuracy gain), while uncertainty quantification achieves excellent calibration (ECE=0.025). The methodology generalizes to earth observation applications requiring fusion of heterogeneous multi-resolution satellite data