Effective variance attention-enhanced diffusion model for crop field aerial image super resolution

Abstract

Image super-resolution (SR) can significantly improve the resolution and quality of aerial imagery. Emerging diffusion models (DM) have shown superior image generation capabilities through multistep refinement. To explore their effectiveness on high-resolution cropland aerial imagery SR, we first built the CropSR dataset, which includes 321,992 samples for self-supervised SR training and two real-matched SR datasets from high-low altitude orthomosaics and fixed-point photography (CropSR-OR/FP) for testing. Inspired by the observed trend of decreasing image variance with higher flight altitude, we developed the Variance-Average-Spatial Attention (VASA). The VASA demonstrated effectiveness across various types of SR models, and we further developed the Efficient VASA-enhanced Diffusion Model (EVADM). To comprehensively and consistently evaluate the quality of SR models, we introduced the Super-resolution Relative Fidelity Index (SRFI), which considers both structural and perceptual similarity. On the × 2 and × 4 real SR datasets, EVADM reduced Fréchet-Inception-Distance (FID) by 14.6 and 8.0, respectively, along with SRFI gains of 27 % and 6 % compared to the baselines. The superior generalization ability of EVADM was further validated using the open Agriculture-Vision dataset. Extensive downstream case studies have demonstrated the high practicality of our SR method, indicating a promising avenue for realistic aerial imagery enhancement and effective downstream applications. The code and dataset for testing are available at https://github.com/HobbitArmy/EVADM.