基于Sentinel-2影像和TPE优化堆叠集成模型估算棉花各生育时期叶片SPAD值估算OA

[Objective]This study aimed to achieve rapid,large-area monitoring of cotton leaf SPAD values using remote sensing technology and to provide technical support for evaluating cotton growth conditions in the Aral reclamation area.[Methods]A total of 22 vegetation indices and color indices were extracted using the Google Earth Engine platform.The mutual information method,recursive feature elimination,and extreme gradient boosting(XGBoost)were separately applied to select spectral features at the seedling,squaring,and flowering and boll-setting stages.The feature-selection performance of the three methods was compared to construct an optimal feature subset for each growth stage and reduce feature redundancy.The tree-structured Parzen estimator(TPE)was used to determine the hyperparameters of K nearest neighbors(KNN),random forest(RF),and categorical boosting(CatBoost).An optimized stacking algorithm(OSA)model was subsequently constructed using the TPE-optimized KNN,RF,and CatBoost models as base learners to improve the accuracy and stability of SPAD value estimation at different cotton growth stages.[Results]When vegetation indices and color indices were jointly used as input variables,the feature combinations selected by XGBoost showed better overall performance.Using the feature subsets selected by XGBoost,the OSA models constructed with TPE-optimized base learners achieved the validation-set coefficients of determination(R2)of 0.805,0.850,and 0.881 at the seedling,squaring,and flowering and boll-setting stages,respectively.And the corresponding mean absolute percentage error(MAPE)values were 0.026,0.022,and 0.019,respectively.The model performed best at the flowering and boll-setting stage,with training-set R2 and MAPE values of 0.917 and 0.008,respectively.[Conclusion]For estimating cotton SPAD values in the Aral Reclamation Area,GRNDVI,GLI,RESAVI,NDVI,EXR,and TVI are recommended as input features at the seedling stage;GNDVI,VARI,GRNDVI,EXR,NDVI,and SI1 at the squaring stage;and OSAVI,LCI,NDI,GBNDVI,NDVI,and SAVI at the flowering and boll-setting stage.At all three growth stages,the OSA model constructed using TPE-optimized KNN,RF,and CatBoost models as base learners is recommended for SPAD value estimation.