Leakage-aware fine-grained crop and treatment classification from UAV imagery

Unmanned aerial vehicle (UAV) studies classifying crop varieties and treatments often report accuracies above 95%, yet trials are organized as plots with constant labels, so random patch splitting places near-duplicate samples in both training and test sets. On a single public trial, MuST-C (17 classes of crop species, genotypes and treatment levels; 80 plots, 13 dates, 3,319 patch sequences) we measure this under otherwise identical conditions: patch-level splitting inflates accuracy by 24.1 points (80.68% versus 56.60%) and macro F1 by 0.295, concentrated on treatment-level classes. Under plot-level cross-validation we benchmark nine published architectures and find that replacing temporal attention with mean pooling gains 5.21 points. We propose HOT-Net, a hierarchical network with a species head, per-species heads and an ordinal head for the monotone nitrogen treatment. HOT-Net is more accurate than each of six published architectures in all four folds, by 4.6 to 14.6 points, with 37–72% fewer parameters; an ablation attributes the gain to the ordinal nitrogen head rather than to the hierarchy itself. With identical 14-channel input including canopy height, its mean accuracy is 5.80 points higher than a selective state-space baseline in 10 of 12 folds, although the corrected difference is not significant (p=0.107).