Cross-platform super-resolution: A diffusion model approach for enhancing satellite imagery with aerial data

• The SR3 model excels on NAIP data but struggles with Planet-NAIP heterogeneity. • NDVI offers more ecological insights than PSNR or SSIM for super-resolution. • Training from scratch outperforms finetuning as the dataset size increases. • Super-resolution on satellite and aerial images is tough due to domain differences. Monitoring Earth’s dynamic systems benefits from high spatial and temporal resolution imagery, a combination rarely available from a single sensor. This study addresses the novel challenge of cross-platform super-resolution (SR), aiming to enhance high-frequency satellite data with high-resolution aerial data. We employed a diffusion-based deep learning model, Super-Resolution via Iterative Refinement (SR3), to upscale 3-m Planet satellite imagery to the 60-cm resolution of the National Agriculture Imagery Program (NAIP) images, a fivefold enhancement. The findings reveal that the “domain gap” between aerial and satellite data is a significant obstacle. While the model performed robustly on single-source data (peak signal-to-noise-ratio, or PSNR, of 27.28 dB), its cross-platform performance was substantially lower (best PSNR of 16.85 dB). Interestingly, models trained from scratch consistently outperformed fine-tuned models, suggesting negative transfer due to the differences between the aerial and satellite data sources. Furthermore, environmental metrics like the Normalized Difference Vegetation Index (NDVI) proved to be more effective performance indicators than standard computer vision metrics (such as PSNR) and structural similarity index measure (SSIM), showing better preservation of critical spectral information for vegetation analysis. This work demonstrates both the potential and the distinct challenges of using diffusion models for super-resolution across different remote sensing platforms. Our findings underscore the importance of tailored approaches in super-resolution and provide insights into leveraging state-of-the-art deep learning techniques for ecological monitoring and resource management.