GeoSR-Bench: New Benchmark Evaluates Super-Resolution for Satellite Imagery
The introduction of a new benchmark dataset, GeoSR-Bench, aims to assess super-resolution (SR) models for remote sensing images, moving past conventional fidelity measures such as PSNR and SSIM. This dataset emphasizes the effectiveness of super-resolved images in applications like land cover classification, biomass estimation, and change detection. GeoSR-Bench includes quality-controlled image pairs that are spatially co-located and temporally aligned from around 36,000 diverse locations. This initiative fills a void in current SR research, which often prioritizes visual fidelity over task-specific outcomes. Detailed in a preprint on arXiv (ID: 2605.00310), the project seeks to enhance the evaluation of SR methods for satellite-based Earth observation in fields like urban planning, agriculture, ecology, and disaster response.
Key facts
- GeoSR-Bench is a downstream task-integrated SR benchmark dataset.
- It evaluates SR models beyond fidelity metrics like PSNR or SSIM.
- The dataset includes about 36,000 locations across diverse land covers.
- Image pairs are spatially co-located, temporally aligned, and quality-controlled.
- Downstream tasks include land cover classification, biomass estimation, and change detection.
- SR techniques are applied to satellite-based Earth observation.
- Applications include urban planning, agriculture, ecology, and disaster response.
- The research is published as arXiv preprint 2605.00310.
Entities
Institutions
- arXiv