Songxi Yang, Tang Sui, Qunying Huang
Super resolution offers a way to harness medium- even low-resolution but historically valuable remote sensing image archives. Generative models, especially diffusion models, have recently been applied to remote sensing super resolution (RSSR), yet several challenges exist. First, diffusion models are effective but require expensive training-from-scratch resources and have slow inference speeds. Second, current methods have limited utilization of auxiliary geospatial information as real-world constraints to reconstruct scientifically realistic images. Finally, most current methods lack evaluation on downstream tasks. In this study, we present a geospatial-aware RSSR framework (GeoAwareDiffSR) that leverages Diffusion Priors and multimodal constraints, trained on a new multi-modal benchmark (LSSR dataset) of paired 30 m Landsat-8 and 10 m Sentinel-2 imagery. Built on a frozen pretrained Stable Diffusion backbone, the proposed GeoAwareDiffSR integrates cross-modal attention with auxiliary knowledge, including Digital Elevation Model (DEM), land cover types, acquisition month, and Synthetic Aperture Radar (SAR) guidance, enhanced by adapters and a tailored Fourier-Normalized Difference Vegetation Index (NDVI) loss to balance spatial details and spectral fidelity. Extensive experiments demonstrate that GeoAwareDiffSR significantly improves crop boundary delineation and recovery, achieving state-of-the-art performance with Peak Signal-to-Noise Ratio/Structural Similarity Index Measure of 32.63/0.84 (RGB) and 23.99/0.78 (IR), and the lowest NDVI Mean Squared Error (0.042), while maintaining efficient inference (0.39 sec/image). Moreover, GeoAwareDiffSR transfers effectively to NASA Harmonized Landsat and Sentinel-2 (HLS) super-resolution, yielding more reliable crop classification (F1: 0.86) than Sentinel-2 (F1: 0.85). These results highlight the potential of RSSR to advance precision agriculture.