Transformer-based and band-selected models for UAV hyperspectral wheat disease classification

Rapid, accurate diagnosis of wheat diseases is essential for targeted crop management and food-security resilience, yet field-scale monitoring remains constrained by labor-intensive visual surveys. Here, we couple unmanned aerial vehicle (UAV) hyperspectral imaging (395–903 nm, 240 bands) with advanced machine learning to form an end-to-end classification pipeline. We first assembled a curated single-pathogen UAV spectral library—3,105 plot-level spectra representing five wheat diseases; bacterial leaf streak, Fusarium head blight and the three major rusts—collected from controlled field trials at US midwestern research stations. Principal-component analysis revealed disease-specific reflectance patterns in both the visible and near-infrared domains, guiding downstream model design. We introduce BandWiseTransformer, a one-dimensional self-attention network tailored to hyperspectral vectors. Automated hyperparameter optimization with Optuna yielded an architecture (128-dimensional embedding, four heads, four encoder layers) that achieved 98 % test accuracy and class-average precision, recall and F1-score of 0.98, decisively outperforming conventional baselines. By analyzing the network’s embedding and classifier weights, we derived a wavelength-importance metric and identified the 10 most informative bands. Random Forest and XGBoost models trained solely on this compressed feature set retained respectable accuracies of 0.86 and 0.88, while slashing dimensionality by 96 % and inference time by two orders of magnitude. SHapley Additive exPlanations (SHAP) analysis confirmed that the selected green, red-edge and near-infrared bands—illustratively 492, 715, 780, 839 and 901 nm—capture biochemical and structural cues linked to pigment loss, red-edge shifts and canopy water status. Thus, transformer attention not only maximises full-spectrum accuracy but also exposes a minimal band subset that underpins lightweight, interpretable classifiers. This study delivers (i) a curated UAV hyperspectral dataset, (ii) a high-accuracy spectral transformer architecture, and (iii) a transparent path from deep-model insights to cost-effective multispectral sensor design. Together, these advances enable real-time scouting of wheat diseases and establish a transferable workflow for hyperspectral crop-health monitoring in diverse pathosystems.