Validation Design Governs Reported Accuracy in Small UAV Hyperspectral Datasets: An Interpretable Feature-Optimization Case Study of Maize Canopy Nitrogen Concentration

Reported accuracy in small unmanned aerial vehicle (UAV) hyperspectral studies depends strongly on how validation is designed, yet leakage-controlled and cross-site protocols are seldom reported. Using 177 plots in three maize fields at the heading stage in Qapqal County, Ili Valley, Xinjiang, China, we quantified this dependence for canopy nitrogen concentration (CNC; mass-based, mg g−1) retrieval from UAV hyperspectral imagery (400–900 nm) with auxiliary LiDAR. A 1829-dimensional feature space served as a realistic testbed. Under a conventional mixed repeated GroupKFold benchmark, the model reached R2 = 0.712, RMSE = 1.023 mg g−1, and RPD = 1.86. Repeating feature ranking inside every training fold lowered this to R2 = 0.612–0.627, and leave-one-field-out validation produced negative R2 for every held-out field. Predictions collapsed toward the training mean with strong slope compression, consistent with a domain shift in which nitrogen level, cultivar, growth micro-stage, and irrigation regime all differ between fields. Feature optimization, not model complexity, drove within-campaign performance: top-140 selection raised R2 from 0.613 to 0.682, whereas ensemble blending and LiDAR (2.7% of TreeSHAP importance) contributed marginally. We report this as a reproducible case study on validation design for small hyperspectral datasets.