Remote Sensing Estimation of Rice Chlorophyll Content Based on UAV Image Feature Selection and PSO-Optimized Ensemble Learning

Chlorophyll content is one crucial indicator of evaluating crop growth and physiological status. Rapid, accurate, and large-scale monitoring of chlorophyll content is vital for the precise diagnosis of crop nutritional status and developmental dynamics. In this study, rice chlorophyll content was estimated using remote sensing data derived from field samples collected under varying nitrogen application levels. The proposed monitoring framework integrated spectral feature selection methods, data augmentation techniques with ensemble learning models, using multi-temporal multispectral imagery acquired by unmanned aerial vehicles (UAVs). Firstly, one two-stage Otsu adaptive thresholding method was utilized to efficiently extract rice pixels from UAV images captured at different growth stages, thereby eliminating non-rice noise in paddy field imagery. Subsequently, spectral vegetation indices, and texture features of rice UAV images were extracted. Three feature selection methods—GRA, mRMR, and SHAP—were employed to identify chlorophyll-sensitive features. Finally, based on Gaussian Mixture Model (GMM)-based data augmentation strategy and the selected sensitive features, one two-layer ensemble learning framework optimized by using Particle Swarm Optimization (PSO) algorithm was proposed to achieve accurate and efficient estimation of rice chlorophyll content across different growth stages. The results demonstrate that the SHAP-based feature selection method consistently achieved the best prediction performance across all growth stages, highlighting its superiority in retaining key feature information and enhancing model stability. Compared with individual machine learning models, the SHAP–PSO-based ensemble learning model exhibited higher accuracy and robustness, with R 2 of 0.641, 0.555, and 0.527, and corresponding rRMSE of 0.055, 0.099, and 0.104 across the different stages. These findings indicate that the proposed ensemble learning framework, which integrates SHAP-based feature selection, GMM-based data augmentation, and PSO optimization, possesses strong generalization capability and stability for the remote sensing estimation of rice chlorophyll content. Moreover, the methodology proposed in this study can serve as a valuable reference for remote sensing-based monitoring of physicochemical parameters in other crops. • A progressive method enables fast extraction of rice pixels via NDCSI and dual Otsu • mRMR, GRA, and SHAP are compared for estimating rice chlorophyll content • A PSO-based ensemble model outperforms 8 methods across 3 rice growth stages • GMM-based data augmentation improves accuracy and stability with small samples