Assessing the Spectral Sufficiency of Wheat Rust Classification: Views and Band Budgets from a UAV Hyperspectral Acquisition

Drone hyperspectral cameras can identify wheat yellow rust but are costly, and labels are scarce. We asked how much spectral information patch-level rust classification needs around grain filling. Using 600 patches of a public ICPR 2026 competition benchmark, we compared the hyperspectral camera’s cube with RGB and five-band multispectral views derived from it, fixing optics, geometry, and timing, and with greedily selected band budgets, under nested cross-validation with Bayesian paired tests and two learners, an automated machine learning search and a tabular foundation model. Five multispectral bands reached a macro-F1 of 0.662 against 0.680 for 101 hyperspectral bands (95 percent credible interval of the difference −0.025 to +0.061, and −0.004 to +0.063 with the foundation model), so neither equivalence nor a relevant gap is excluded, and no greedily selected bands beat them. The distributed RGB images lost 0.117 macro-F1 against the same cube bands, at least as much as omitting the red-edge and near-infrared bands, and an exploratory control traced this loss to their per-image contrast stretch and quantisation. In exploratory analyses within vegetated canopy, the cube kept a Health versus Rust advantage of 0.056 in the area under the receiver operating characteristic curve (−0.002 to +0.113), also with canopy greenness balanced, but no wavelength window stood out, and no configuration separated rust from healthy canopy reliably.