HyperBench reveals fragility in hyperspectral super-resolution evaluations
New framework tests HSR models on 70 configurations, exposing hidden weaknesses
Hyperspectral super-resolution (HSR) reconstructs high-spatial-resolution images by fusing low-resolution hyperspectral with high-resolution multispectral data. Current evaluation relies on synthetic experiments using Wald’s protocol, but implementations vary widely—often using just one PSF, one or two SRFs, and a couple of downsampling factors. This makes reported performance figures hard to compare and reproduce.
HyperBench addresses this by standardizing degradation configurations: 10 PSFs, 4 SRFs from operational sensors (e.g., Sentinel-2), configurable spatial downsampling, and additive white Gaussian noise. The framework automates large-scale evaluation and structured logging, decoupling model development from experimental design. Evaluating six recent methods across a 70-configuration sweep on four scenes, the authors found inter-method PSNR spread widens from ~5 dB on the easiest PSF to over 13 dB on the hardest—a fragility structurally invisible to conventional single-configuration protocols. HyperBench code is publicly available.
- Supports 10 point spread functions (PSFs) and 4 spectral response functions (SRFs) from real multispectral sensors
- Evaluated six HSR methods across 70 different degradation configurations on four hyperspectral scenes
- PSNR spread between methods widens from ~5 dB on easiest PSF to >13 dB on hardest, exposing evaluation fragility
Why It Matters
Standardized evaluation prevents overfitting to easy conditions, vital for remote sensing, medical imaging, and precision agriculture.