Preserving spectral fidelity in hyperspectral super-resolution with gradient-weighted loss
Abstract
Deep learning has substantially improved the spatial enhancement of hyperspectral imagery, yet training objectives remain dominated by band-agnostic pixel-wise fidelity terms that treat every pixel as equally informative, regardless of how much diagnostic spectral structure its spectrum contains. A super-resolved image can therefore achieve favourable global reconstruction scores while distorting the narrow reflectance features that separate one urban surface material from another, a limitation for infrastructure monitoring, where reliable spectral discrimination between materials such as asphalt, bitumen and bare soil is essential. This paper proposes a gradient-weighted spectral angle (GWSA) loss that re-weights the spectral-angle penalty of each spectrum by a factor derived from the spectral-gradient magnitude of the high-resolution reference, so that pixels whose spectra contain steep band-to-band reflectance transitions receive greater emphasis. Experiments on the Pavia University and Pavia Centre benchmarks (ROSIS sensor, 1.3 m GSD) use a lightweight band-grouped 3D residual network for ×2 super-resolution under idealised bicubic downsampling and a sensor-realistic pipeline of Gaussian point-spread-function blur, spatial decimation, additive noise and bicubic resampling. Relative to an MSE-only objective on the same network, GWSA reduces the mean spectral angle on the strictly unseen test region of each scene by 5–13% across the four scene–degradation conditions and the asphalt-class spectral angle by up to 24%, at no consistent PSNR cost. Relative to a uniform spectral-angle loss, the difference is small and condition-dependent. The gradient-weighted objective, however, concentrates its gains in the high-gradient regions it up-weights and achieves the highest downstream classification accuracy of the compared losses under sensor-realistic degradation; class-mean asphalt spectra confirm that both spectral losses remove the large near-infrared errors of pixel-wise training. Weighting the spectral penalty by the gradient structure of the reference therefore improves material-level reconstruction fidelity in urban hyperspectral scenes at negligible spatial cost.