Novel Machine Learning Framework for High-Resolution Sorghum Biomass Estimation Using Multi-Temporal UAV Imagery

Abstract

Accurate assessment of crop health and yield potential facilitates precise estimation of above-ground biomass (AGB). Traditional AGB estimation methods are often limited by their destructive, labor-intensive nature. This study developed a rapid, non-destructive approach to estimate sorghum AGB using high-resolution unmanned aerial vehicle (UAV) data and machine learning (ML). A two-year field experiment tested four irrigation strategies: full irrigation at 100% of crop evapotranspiration (S1), partial deficits at 75% and 50% of S1, and a rain-fed (S4). This gradient assessed the ML model's robustness across diverse conditions, yielding 216 AGB measurements reflecting variable plant responses to water stress. Multispectral and canopy height data were derived from UAV imagery collected during the sorghum growing season. Three ML algorithms-Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbors (K-NN)-were applied. RF outperformed others with an R2 of 0.80, RMSE of 0.78 kg m(-)2, and MAE of 0.58 kg m(-)2, followed by SVM (R2 = 0.64, RMSE = 1.08 kg m(-)2, MAE = 0.77 kg m(-)2), while K-NN showed the lowest accuracy (R2 = 0.50, RMSE = 1.26 kg m(-)2, MAE = 0.96 kg m(-)2). Optimal RF hyperparameters were identified, and estimated AGB aligned closely with ground measurements, showing no significant differences. Spatial AGB maps effectively highlighted variability across treatments. This study demonstrates that UAV-based remote sensing combined with ML offers a reliable, non-destructive method for sorghum AGB estimation, enhancing precision agriculture applications such as irrigation management, crop monitoring, and yield prediction.