Non-destructive multinutrient monitoring in Bahiagrass pastures from UAV multispectral imagery using a machine-learning framework

This study developed and evaluated a non-destructive unmanned aerial vehicle (UAV)-based framework for predicting plant nutrient status in bahiagrass pastures using machine learning. Leaf tissue samples and UAV multispectral imagery were collected approximately monthly across 20 monitoring rounds spanning a continuous ∼20-month field experiment at three commercial ranches and analyzed for 11 tissue nutrients: nitrogen (N), phosphorus (P), potassium (K), magnesium (Mg), calcium (Ca), sulfur (S), boron (B), zinc (Zn), manganese (Mn), iron (Fe), and copper (Cu). This represents one of the longest continuously evaluated UAV–nutrient modeling datasets in pasture systems, enabling robust assessment of seasonal and interannual variability. UAV imagery was processed into spectral predictors extracted from 10-m central buffers within each experimental unit, including four spectral bands and 12 vegetation indices summarized by mean, standard deviation, and coefficient of variation. Seasonal context was incorporated through cyclic month predictors. Target-specific Random Forest models were optimized using nested spatial-unit-blocked cross-validation, and predictive performance was summarized from outer-fold out-of-fold (OOF) predictions using the coefficient of determination (R²), root mean square error (RMSE), and mean absolute percentage error (MAPE). The highest out-of-fold performance was obtained for nitrogen (R² = 0.692, RMSE = 0.261, MAPE = 22.11%), phosphorus (R² = 0.628, RMSE = 0.050, MAPE = 17.14%), potassium (R² = 0.699, RMSE = 0.236, MAPE = 18.48%), manganese (R² = 0.570, RMSE = 37.023, MAPE = 65.96%), and sulfur (R² = 0.554, RMSE = 0.045, MAPE = 18.15%). The remaining nutrients showed lower explained variance, with R² values of 0.521 for Cu, 0.423 for Zn, 0.369 for Ca, 0.328 for Mg, 0.209 for B, and 0.183 for Fe. Predictor-relevance analyses indicated that visible-band canopy contrast, cyclic seasonal covariates, and within-unit spectral heterogeneity contributed differently across nutrient targets. These results show that UAV multispectral imagery can support non-destructive estimation of selected tissue nutrients in bahiagrass pastures, particularly N, P, K, S, and Mn. The framework provides a scalable basis for georeferenced nutrient monitoring in pasture systems and establishes a transparent distinction between stronger nutrient targets and weaker exploratory ones. Importantly, the long-term 20-month experimental design enhances the robustness and generalizability of the proposed framework under realistic seasonal and management variability.