Abstract: Mixed pixels containing cotton canopy, soil, and shadow in unmanned aerial vehicle (UAV) remote sensing imagery can reduce the accuracy of cotton canopy flavonol content (Flav) inversion, while different machine-learning models may show substantial differences in estimation performance. To address these issues, a UAV-based Flav inversion framework integrating canopy-mask constraints, mixed-pixel decomposition, multi-type feature fusion, and model optimization was developed. Visible-light and multispectral UAV images were acquired at flight altitudes of 50, 80, and 110 m, together with synchronous ground measurements of Flav. Under identical mask constraints, endmember sets, and high-purity pixel screening conditions, four mixed-pixel decomposition methods were compared: the linear spectral mixing model (LSMM), generalized bilinear model (GBM), L2-regularized generalized bilinear model (L2-GBM), and LSMM-GBM hybrid model (Hybrid). LSMM produced mean cotton endmember abundances of 0.9961-0.9973 across the three flight altitudes, with standard deviations of 0.0048–0.0061, showing better abundance stability than the other models. After LSMM processing, 16 of the 18 vegetation indices showed increased absolute Pearson correlation coefficients with Flav at all three flight altitudes. The mean increases were 0.0635, 0.0714, and 0.0715 at 50, 80, and 110 m, respectively, indicating that mixed-pixel decomposition effectively enhanced the spectral response to Flav. Pearson correlation analysis and collinearity diagnostics were then used to select six vegetation indices, two color features, and two texture features. Seven candidate feature sets were constructed from individual feature types and their combinations to evaluate the contribution of spectral, color, and texture information to Flav estimation. Six regression models were evaluated, 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). Key hyperparameters were optimized by cross-validation within the training set. On the basis of PLSR optimization, Bagging, correlation-weighted adaptive (CWA), and differential evolution (DE) strategies were further compared to assess whether additional optimization could improve predictive performance. Considering estimation accuracy, model stability, and complexity, the cross-validation-optimized PLSR model (CV-PLSR) was selected as the final model. It achieved a coefficient of determination (R²) of 0.9433, a root mean square error (RMSE) of 0.0339, and a mean absolute error (MAE) of 0.0266 on the test set, indicating a favorable balance between prediction accuracy and model complexity. To evaluate model transferability, independent data collected at the boll-opening stage in 2025, the flowering-boll stage in 2026, and the boll-opening stage in 2026 were used for cross-field, cross-growth-stage, and cross-year validation. During validation, the selected features, preprocessing procedure, model structure, and parameters were kept fixed, and no feature reselection, model recalibration, or parameter re-optimization was performed. Across the three independent datasets, CV-PLSR achieved R² values of 0.8401-0.9100, RMSE values of 0.0289-0.0363, and MAE values of 0.0228-0.0296. Although model performance declined as the validation conditions differed more from the model-development dataset, the estimation errors remained relatively low. These results indicate that the proposed model has a certain degree of cross-temporal, cross-field, and cross-growth-stage stability within the same ecological region. The proposed framework provides technical support for rapid and non-destructive field monitoring of cotton flavonol status and for auxiliary analysis of physiological differences among cotton breeding materials. The results further confirm that background suppression and multi-feature integration can improve the robustness of UAV-based physiological trait estimation under changing field conditions.