UAV-based estimation of sunflower leaf flavonol content across multiple flight heights using RGB-multispectral canopy...

Rapid and non-destructive estimation of leaf flavonol content (Flav) is important for assessing crop physiological status and stress response. In this study, unmanned aerial vehicle (UAV)-based red, green, and blue (RGB) and multispectral imagery was used to estimate sunflower leaf Flav across the seedling, budding, flowering, and maturity stages. Data were collected at a single experimental site in the Yellow River Delta, China, during the 2025 growing season. UAV images were acquired at flight heights of 30 m, 50 m, and 80 m, and ground measurements of Flav were obtained within the same acquisition window using an MPM-100 plant multispectral pigment meter. A total of 320 plot-level observations were obtained from 80 sampling plots, which were divided at the plot level into 56 training plots and 24 testing plots, with all four growth-stage observations from the same plot assigned to the same subset. Feature selection and model development were conducted using the training data, which were further divided for hyperparameter optimization, whereas the testing set was used for comparative performance evaluation. A total of 44 UAV-derived features, including 22 vegetation indices, 16 texture features, and six color features, were extracted from sunflower canopy imagery. Pearson correlation coefficient (PCC) analysis and the Boruta algorithm were combined to identify compact feature subsets, and six regression models, including K-nearest neighbors, random forest, CatBoost, support vector regression, LightGBM, and Transformer regression, were evaluated. Genetic algorithm (GA), particle swarm optimization (PSO), and grey wolf optimizer (GWO) were further used to tune the Transformer model at the intermediate flight height of 50 m. The PCC-Boruta strategy reduced feature dimensionality and produced favorable predictive performance on the testing set. Among the evaluated regression models, the Transformer model achieved the best overall performance. Using the optimal feature combination, the PCC-Boruta-Transformer model obtained test-set R² values of 0.932, 0.911, and 0.909 at 30 m, 50 m, and 80 m, respectively, with corresponding root mean square error (RMSE) values of 0.060, 0.069, and 0.070. Thus, the 30 m model provided the highest predictive accuracy among the three unoptimized flight-height models. The 50 m dataset, representing the intermediate level among the three evaluated flight heights, was used for subsequent hyperparameter optimization. Among the evaluated optimization configurations, the GWO-tuned Transformer achieved the best observed testing-set performance, with an R² of 0.947 and an RMSE of 0.053. These results demonstrate the feasibility of the evaluated workflow under the specific site, growing-season, sensor, and management conditions of this experiment.