Inversion of cotton flavonol content based on UAV imagery and machine learning optimization

Effective canopy spectral information is confined to mixed pixels containing cotton canopy, soil, and shadow components in unmanned aerial vehicle (UAV) remote sensing imagery, leading to low inversion accuracy of cotton flavonol content (Flav). There are also various performances among different machine learning models and feature combinations. In this study, an inversion framework was developed for cotton Flav using UAV remote sensing. A series of optimizations were integrated with canopy-mask constraints, mixed-pixel decomposition, multi-type feature fusion, and machine learning. UAV visible-light and multispectral images were acquired at three flight altitudes of 50, 80, and 110 m, while cotton leaf Flav was synchronously measured in the field. Four mixed-pixel decomposition methods were compared for multispectral imagery, including the linear spectral mixing model (LSMM), the generalized bilinear model (GBM) incorporating interactions among endmembers, the L2-regularized generalized bilinear model (L2-GBM), and the LSMM-GBM hybrid model (Hybrid), under canopy-mask constraints, endmember sets, and high-purity pixel selection. A total of 10 visible-light vegetation indices, 8 multispectral vegetation indices, 9 color features, and 8 texture features were extracted subsequently. Pearson correlation analysis with collinearity diagnostics was used for feature screening. 6 vegetation indices, 2 color features, and 2 texture features were retained as sensitive variables for cotton Flav estimation. Seven feature combinations were constructed, including vegetation indices, color features, texture features, vegetation indices with color features, vegetation indices with texture features, color features with texture features, and the combination of vegetation indices, color features, and texture features. Six regression algorithms were employed to establish cotton Flav estimation models, including partial least squares regression (PLSR), random forest (RF), support vector regression (SVR), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and categorical boosting (CatBoost). The parameters of six algorithms were tuned using internal cross-validation in the training dataset during optimization, according to the structural characteristics of each algorithm. Furthermore, a comparison was made on PLSR parameter tuning with three optimization strategies, namely bagging ensemble learning, correlation-weighted adaptive optimization, and differential evolution (DE). The coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) of the test dataset were evaluated to determine the final inversion model. After that, the data were collected at the boll-opening stage in 2025, the flowering-boll stage in 2026, and the boll-opening stage in 2026. Cross-field, cross-growth-stage, and cross-year generalization tests were conducted to evaluate the generalization of the model under varying field, phenological, and annual conditions. The results showed that the mean cotton endmember abundance ranged from 0.9961 to 0.9973 at flight altitudes of 50, 80, and 110 m, with standard deviations from 0.0048 to 0.0061. The cotton endmember abundance after estimation was more stable than that of the GBM, L2-GBM, and Hybrid. There were strong correlations between the 18 vegetation indices extracted from the multispectral imagery and Flav after LSMM processing. Among them, 16 vegetation indices showed that the absolute Pearson correlation coefficients increased with Flav at all three flight altitudes. The average increase in the absolute correlation coefficients was 0.0635, 0.0714, and 0.0715 at 50, 80, and 110 m, respectively, indicating that mixed-pixel processing improved the response of vegetation indices to variations in cotton Flav. Among the six machine learning models, PLSR achieved the highest overall accuracy for Flav estimation. After parameter optimization, the cross-validation-tuned PLSR model (CV-PLSR) achieved an R² of 0.9433, an RMSE of 0.0339, and an MAE of 0.0266. The generalization tests with data from different growth stages and years showed that the R² of CV-PLSR ranged from 0.8401 to 0.9100, while RMSE ranged from 0.0289 to 0.0363 and MAE ranged from 0.0228 to 0.0296. Stable estimation performance was achieved in the field, growth stages, and years. The findings can provide technical support for rapid and non-destructive field monitoring of cotton flavonol content and physiological differentiation of cotton breeding materials.