GeoSR-Bench Reveals Sharper Satellite AI Images Often Hurt Downstream Tasks
GeoSR-Bench proves that sharper satellite images don't automatically mean better AI predictions — and introduces a benchmark to fix that
Takeaways
- GeoSR-Bench is the first SR benchmark for satellite imagery that evaluates models on real downstream tasks, not just pixel-level quality metrics.
- The dataset covers ~36,000 locations globally with real paired imagery spanning 500m (MODIS) to 0.6m (NAIP) resolution.
- 270 experiments across 9 SR models (GAN, diffusion, transformer, neural operator) and 5 downstream tasks show PSNR/SSIM gains often don't translate to better task performance.
- Correlations between SR metrics and downstream task performance can be negative — meaning higher PSNR can mean worse segmentation or biomass estimation.
- Two real-world SR tracks are covered: MODIS→Landsat-8 (500m→30m) and Sentinel-2→NAIP (10m→sub-meter).
- Paper and benchmark available at arxiv.org/abs/2605.00310.
Super-resolution for satellite imagery has been on a tear. GAN models, diffusion models, transformer-based architectures , all of them can take a blurry 500-meter MODIS pixel and reconstruct something that looks strikingly like a 30-meter Landsat image. The numbers look great: PSNR up, SSIM up, paper accepted. But a new benchmark called GeoSR-Bench asks a harder question: does that sharper image actually help you map forests, detect roads, or estimate crop yields any better? The answer, it turns out, is often no , and sometimes the opposite.
The metric trap in satellite AI
Super-resolution has been increasingly developed for satellite-based Earth observation, with applications in urban planning, agriculture, ecology, and disaster response. The promise is clear: if you can upscale cheap, freely available imagery from MODIS or Sentinel-2 to the resolution of expensive commercial satellites, you unlock a huge amount of analytical potential without the cost.
However, existing SR studies and benchmarks typically use fidelity metrics such as PSNR or SSIM, whereas the true utility of super-resolved images lies in supporting downstream tasks such as land cover classification, biomass estimation, and change detection. PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index) measure how closely a reconstructed image matches a reference pixel-by-pixel. They say nothing about whether the reconstructed image is actually useful for a segmentation model trying to identify wetlands or a regression model estimating canopy height.
Conventional metrics like PSNR and SSIM poorly correlate with task performance , a finding corroborated in ground-level pollution mapping, where feature preservation outweighed perceptual quality. GeoSR-Bench turns this suspicion into a rigorous, large-scale empirical result.
What GeoSR-Bench actually is
GeoSR-Bench is a downstream task-integrated SR benchmark dataset built to evaluate SR models beyond fidelity metrics. It comprises spatially co-located, temporally aligned, and quality-controlled image pairs from about 36,000 locations across diverse land covers, spanning resolutions from 500m to 0.6m. That span , from coarse MODIS pixels down to sub-meter aerial imagery , covers the full range of practically useful cross-platform SR scenarios.