Estimation of leaf nitrogen content in dryland forage maize using UAV-based hyperspectral imaging and machine learning

This study integrates UAV-based hyperspectral imaging with ensemble learning to identify an optimal spectral estimation model for predicting leaf nitrogen content (LNC) in dryland forage maize, providing a methodological reference for improving production efficiency and quality. The study was conducted on the Loess Plateau in central Gansu province, China, with forage maize as the target crop. Hyperspectral data were acquired using a V185G integrated gimbal hyperspectral imaging system mounted on a UAV. Spectral indices were generated from all possible two-band combinations using original reflectance spectra, first-derivative spectra, and continuum-removed spectra. Six machine-learning algorithms were evaluated, and Voting and Stacking ensemble models were further developed to select the best-performing approach. Transformed spectra substantially strengthened the relationships between spectral indices and LNC compared with the original bands. Among the six individual models, random forest regression (RFR), <italic>K</italic>-nearest neighbors (KNN), XGBoost, and gradient boosting decision tree (GBDT) achieved relatively high accuracy across maize growth stages, with test-set <italic>R</italic>² values of 0.7165-0.7713 and RMSE values of 2.4265-2.8296. These four models were then combined to build the ensemble models, both of which achieved test-set<italic> R</italic>² &gt; 0.7459 and RMSE &lt; 2.6358. The Voting ensemble based on first-derivative spectra delivered the best performance (<italic>R</italic>² = 0.8152, RMSE = 2.1253), indicating improved predictive accuracy and robustness through model integration. Overall, the Voting-first-derivative spectra (Voting-FDS) model enables rapid estimation of LNC at key growth stages, supporting in-season nutrient management and high-quality production in dryland forage maize.