Adaptive multi-scale feature aggregation transformer network for single remote sensing image super-resolution

Remote sensing image super-resolution (RSISR) plays a key role in recovering spatial details and improving image quality from satellite imagery. In recent years, transformer-based methods have shown excellent performance in RSISR tasks. Despite the higher computational efficiency of local self-attention calculations compared to global self-attention calculations, its limited receptive field restricts the model from effectively modeling the complex scale diversity and long-range dependencies of ground observation targets. Moreover, the intermediate features of existing methods contain blocking artifacts, leading to different degrees of feature edge distortion and texture detail loss. To address the above issues, this paper proposes the adaptive multi-scale feature aggregation transformer network (AMFAT), which improves the feature representation capability through dynamic weighting and cross-window interaction. Specifically, the adaptive context channel attention (ACCA) is designed to fuse multi-branch features using dynamic weights for object-guided context adaptation. In addition, the mixed-scale token attention (MSTA) is constructed to eliminate blocking artifacts through cross-window interaction. Meanwhile, simple gating units with spatial enhancement operations are introduced into the feed-forward network (FFN) to optimize local feature aggregation. We conducted extensive experiments on four publicly available remote sensing datasets, and the results show that, compared to other methods, AMFAT exhibits excellent performance and adaptability both in terms of quantitative metrics and visual quality. The model will be available at https://github.com/sq-3768/AMFAT.