OFTNet: Omni-Feature and Texture-Aware Enhancement Learning Network for Continuous Remote Sensing Image Super-Resolution

Continuous-scale super-resolution (CSSR) for remote sensing images (RSIs) offers critical flexibility for diverse Earth observation tasks, overcoming the fixed-scale limitations of conventional SR methods. However, current CSSR methods are primarily tailored for natural images, failing to address the unique characteristics of remote sensing scenes. Unlike object-centric natural images with clear foregrounds, RSIs feature arbitrary orientations, vast scale variations, and dense repetitive patterns. First, standard hierarchical extractors in existing methods exhibit limited receptive fields, struggling to capture the global context required for large-scale land covers while preserving the fine-grained details of small targets. Second, the inherent spectral bias in implicit neural representations exacerbates the loss of high-frequency information, causing severe over-smoothing in regions with intricate textures. To address these challenges, we propose OFTNet (Omni-Feature and Texture-aware enhancement learning Network), a novel network specifically designed for continuous-scale RSI super-resolution. Our method introduces a powerful omni-feature enhancement and refinement pipeline, comprising three key modules: an Omni-Feature Enhancement Module (OFEM) that synergistically models global spatial dependencies, inter-channel correlations, and multi-scale local patterns; a Texture-Aware Enhancement Module (TAEM) that explicitly preserves high-frequency information and models non-local spatial correlations; and a Learnable Weighting Aggregation Module (LWAM) for intelligent, context-aware feature aggregation. Extensive experiments on commonly-used benchmark remote sensing datasets demonstrate that OFTNet achieves state-of-the-art performance across a wide range of in-distribution scales, while also showing robust and highly competitive performance on out-of-distribution scales.