Dynamic Patch-Based Inference for High-Resolution Remote Sensing Image Super-Resolution

Remote sensing image super-resolution (SR) aims to restore sharp edges and fine texture details to improve the reliability of image interpretation and target recognition. However, applying uniform computational effort to an entire large-scale image is inefficient, since structurally complex regions require stronger reconstruction capability, whereas smooth background regions often suffer from redundant computation. To address this problem, this paper proposes The proposed method, a dynamic patch-based inference framework for high-resolution remote sensing image SR under computational constraints. The framework first adopts a quadtree-based adaptive partitioning strategy driven by local structural responses to generate content-aware variable-size patches. Then, a lightweight/high-performance dual-expert reconstruction scheme is employed, and an upgraded routing strategy allocates limited high-performance computation to regions with higher reconstruction benefit. To support large-scale inference with irregular patches, context-extended packaging, bucket-based batching, and Gaussian-weighted backfilling are further introduced to reduce boundary artifacts and maintain full-image spatial continuity. Experiments on the NAIP remote sensing dataset demonstrate that the proposed framework achieves a favorable trade-off between reconstruction quality and computational cost. Compared with uniform lightweight and uniform high-performance inference, The proposed method improves reconstruction quality with substantially lower additional FLOPs, while ablation results verify the effectiveness of adaptive partitioning, routing, context extension, and weighted fusion. These results indicate that the proposed method is suitable for resource-constrained remote sensing SR deployment.