Cross-domain continuous-scale remote sensing image super-resolution via meta-weight learning

Continuous-scale Remote Sensing Image Super-Resolution offers significant flexibility but relies critically on spatial anchor feature aggregation. Prevailing methods utilize fixed geometry-based weighting, limiting detail recovery and causing substantial performance degradation in cross-domain scenarios where natural image pre-trained models fail to adapt. To address this, we introduce Meta-Weight Learning, a meta-learning approach featuring a novel implicit decoder that dynamically predicts content-aware weights for each anchor. These weights integrate feature representations and relative geometric positions, enabling adaptive fusion without handcrafted rules. This strategy provides superior spatial adaptability, effectively bridging the cross-domain gap without requiring encoder fine-tuning. Comprehensive experiments validate the efficacy of our proposed framework, demonstrating state-of-the-art performance and outperforming existing methods in both visual quality and quantitative metrics for continuous-scale RSISR across domains.