遥感探测小麦条锈病严重度的GBRT模型研究

为了提高小样本数据模型的稳定性，构建具有更高精度和鲁棒性的小麦条锈病遥感探测模型。首先基于辐亮度和反射率荧光指数方法提取了冠层日光诱导叶绿素荧光（SIF）数据，然后结合对小麦条锈病病情严重度敏感的反射率光谱指数并基于改进的分类与回归树（CART）——梯度提升回归树（GBRT）算法，构建了融合反射率和冠层SIF数据的小麦条锈病遥感探测的GBRT模型，并将其与CART及多元线性回归（MLR）模型进行比较。结果表明：①反射率导数荧光指数D705/D722、短波红外谷反射率和反射率比值荧光指数R740/R800是影响遥感探测小麦条锈病严重度精度的主控因素，其中SIF数据的重要性值高于反射率光谱数据，冠层SIF能够比反射率光谱更加敏感地反映小麦条锈病害信息；②GBRT模型病情指数（DI）预测值和实测值间的均方根误差（RMSE）比CART和MLR模型分别减小了15.50%和13.49%，决定系数（R²）分别提高了6.16%和11.57%，GBRT模型估测DI值更接近于实测值，且估测结果波动小，鲁棒性高；CART模型在小样本数据中易将不同特征的数据集划分为同一特征的子集，预测结果波动较大；MLR模型的预测结果相对比较稳定，但其预测精度较低。

GBRT Model for Detecting the Severity of Wheat Stripe Rust by Remote Sensing

In order to improve the stability of the small sample data model， a remote sensing detection model of wheat stripe rust with higher accuracy and better robustness was constructed. Firstly， the data of canopy solar-Induced chlorophyll Fluorescence （SIF） were extracted based on radiance and reflectance fluorescence index method， and then combined with reflectance spectral index sensitive to severity of wheat stripe rust， the Gradient Boost Regression Tree （GBRT） was used to detect wheat stripe rust. By comparing GBRT model with CART and Multiple Linear Regression （MLR） model， the results showed that： （1） Reflectivity derivative fluorescence index D705/D722， short-wave infrared Valley reflectance and reflectance ratio fluorescence index R740/R800 were the main factors affecting the accuracy of remote sensing detection of wheat stripe rust. The importance of chlorophyll fluorescence data was higher than that of reflectance spectrum data， and canopy SIF could reflect wheat stripe rust information more sensitively than reflectance spectrum. （2） Compared with CART model and MLR model， the Root Mean Square Error （RMSE） of GBRT model was reduced by 15.50% and 13.49%， and the determination coefficient （R²） was increased by 6.16% and 11.57% respectively. The estimated DI value of GBRT model is closer to the measured value， and the fluctuation of the estimated result is low， and the robustness of CART model is high. In small sample data， it is easy to divide data sets with different features into subsets of the same feature， and the prediction results fluctuate greatly. The prediction results of MLR model are relatively stable， but its prediction accuracy is low.