HAIP Blackbird v2 Hyperspectral Dataset and Classification Pipeline for Winter Oat Discrimination

Reproducibility package for the article "Classifying Winter Oat Varieties and Foliar Nutrient Treatments Using UAV Hyperspectral Imaging and Machine Learning" (Precision Agriculture, 2025). Contents: - Hyperspectral reflectance dataset: 686,000 pixels × 100 spectral bands (500–1000 nm, 5 nm resolution) acquired with a HAIP Blackbird v2 frame camera (DJI Matrice 300 RTK, 60 m AGL) over winter oat experimental plots at the University of Debrecen, Hungary. Eight varieties × four foliar treatments (control, Si, S, Si+S) = 32 classes. - Python classification pipeline: 13 machine learning classifiers (XGBoost, Random Forest, SVM, LightGBM, AdaBoost, Extra Trees, Gradient Boosting, k-NN, SGD, LDA, QDA, Naive Bayes, Neural Network) evaluated across three preprocessing methods (raw reflectance, PCA, MNF). Includes hyperparameter tuning (RandomizedSearchCV), 10-fold stratified cross-validation, and extended GPU neural network evaluation (8 architectures including ResNet MLP, 1D-CNN, Attention). - SHAP interpretability analysis: feature importance ranking by spectral wavelength, per-class decomposition, and SHAP-guided band selection. - Statistical tests: Friedman rank test with Nemenyi post-hoc, Bonferroni-corrected paired t-tests. - Additional analyses: spatial cross-validation (plot-level vs pixel-level), learning curves, vegetation index comparison, stacking ensemble, two-stage hierarchical classification, prediction confidence, computational cost benchmarks. - All pre-computed results (JSON, CSV, PNG) for direct inspection without re-running. Key results: XGBoost with PCA achieved 94.88% accuracy for variety discrimination (8 classes), 92.53% for treatments (4 classes, AdaBoost), and 89.51% for 32-class variety–treatment combinations. SHAP analysis identified the green-onset region (500–515 nm) as most discriminative for varieties and NIR water-absorption (930 nm) for treatments.