High-performance extraction method of highland barley crops utilizing feature optimization and machine learning algorithms based on Sentinel-2 images at key phenological stages: a case study of Linzhou County

Highland barley remote sensing faces challenges such as high feature redundancy and unstable accuracy. In this study, five phenological Sentinel-2 images were selected, and 26 features (spectral, texture, vegetation, water and red-edge indices) were extracted and optimized via PCA and RF-RFE. Five classification algorithms (RF, SVM, CNN, CatBoost and XGBoost) were compared. Results show that after feature optimization, ensemble learning models (XGBoost, CatBoost and RF) maintained overall accuracies above 85% across all stages, with the flowering and milk ripening stages yielding the best performance. XGBoost demonstrated the strongest stability, achieving overall accuracies of 94.01% and 95.49% at flowering and milk ripening, respectively. Furthermore, PCA-derived spectral features, red-edge indices and some texture features exhibited good universality across different phenological stages, with PCA_1, PCA_2, PCA_3, NDVIre1 and Mean identified as key feature variables for highland barley extraction.