Optimized Machine Learning Algorithms to Enhance Rice Protein Content Monitoring with Multispectral UAV Imagery

초록

Purpose Accurate estimation of grain protein content (GPC) in rice using remote sensing is essential for evaluating grain quality yet remains challenging due to limited spectral sensitivity and variability in multispectral data. This study developed a predictive framework for estimating GPC at harvest using unmanned aerial vehicle (UAV)-based multispectral imagery, while accounting for factors such as leaf senescence, canopy structure, and panicle development during the ripening stage.Methods A fixed-wing UAV equipped with red, green, and near-infrared (NIR) sensors was used to acquire field data. Advanced machine learning models were implemented and optimized using Bayesian optimization and the asynchronous successive halving algorithm (ASHA) to automate hyperparameter tuning. In addition, hybrid optimization approaches combining genetic algorithm (GA) and particle swarm optimization (PSO) were evaluated for improving fuzzy tree (FT) model performance.Results It is found that both Bayesian and ASHA optimization significantly enhanced prediction accuracy, with ASHA demonstrating faster convergence. Bayesian optimization achieved the best overall performance (R = 0.896, IOA = 0.939, NSE = 0.795, RMSE = 0.107), while ASHA produced comparable results (R = 0.892, IOA = 0.936, NSE = 0.786, RMSE = 0.111). In contrast, GA- and PSO-based hybrid models showed moderate accuracy, with GA-GA and PSO-PSO configurations yielding higher errors.Conclusion Overall, the findings demonstrate that advanced optimization techniques, particularly Bayesian and ASHA, substantially improve GPC prediction using UAV-based multispectral data, highlighting their potential for precision agriculture and informed crop management.

키워드