基于无人机多光谱影像的土壤盐分反演模型

为探究不同作物覆盖下不同深度的土壤盐分快速反演模型，该研究采集苜蓿、玉米覆盖下0～15、>15～30、>30～50 cm层深度的土壤盐分含量，基于无人机多光谱影像数据，提取各地块采样点的光谱反射率，在此基础上引入红边波段计算光谱指数作为特征变量，采用支持向量机递归特征消除算法（Support Vector Machine-Recursive Feature Elimination，SVM-RFE）以筛选光谱指数及未经过筛选的全指数组作为模型输入组，共构建出36个基于随机森林（Random Forest，RF）、极限学习机（Extreme Learning Machine，ELM）、BP神经网络（Back Propagation Neural Network）等机器学习模型，确定不同作物覆盖下的最佳土壤盐分反演模型。结果表明：SVM-RFE算法筛选光谱指数构建模型精度优于未进行筛选构建的模型。对于苜蓿和玉米覆盖土壤，整体上，RF反演效果优于ELM模型和BPNN模型，反演结果能体现真实土壤盐分含量，在0～15和>30～50 cm土层上，RF模型反演效果优于其他模型，苜蓿样地验证集决定系数Rp2分别为0.71、0.58，验证集均方根误差RMSEp分别为0.026、0.033，玉米样地Rp2分别为0.67、0.64，RMSEp分别为0.111、0.094，在>15～30 cm土层上ELM反演效果较好，苜蓿样地Rp2为0.58，RMSEp为0.039，玉米样地Rp2为0.68，RMSEp为0.059。0～15 cm是作物覆盖下的土壤含盐量最佳反演深度，验证集平均决定系数R2为0.65，均方根误差RMSE为0.084。研究结果可为土壤盐分的快速反演提供理论依据。

Soil salinity inversion model based on the multispectral images of UAV

Abstract: Soil salinization seriously affects the growth and yield of crops and has threatened food security in China.Timely acquisition of soil salinity information is the key to salinization control. The Taolai River Basin is an important agricultural planting area in the inland area of northwest China. The saline-alkali land in the basin is widely distributed, and the Bianwan Farm is a representative area of soil salinization. This study takes the Bianwan Farm in the Taolai River Basin as the sampling area. The area is located in Suzhou District, Jiuquan City, Gansu Province. In order to explore the rapid inversion model of soil salinity at different depths under different crop cover, this paper mainly collected the soil salinity values at different depths of 0-15 cm, 15-30 cm and 30-50 cm under the cover of alfalfa and corn in the phenological period of the Bianwan Farm, and obtained the multi-spectral image data of the UAV at the same time. By extracting the spectral band reflectance of different plot acquisition points, and on this basis, the red edge band is introduced to calculate the spectral index (the red edge band can effectively improve the inversion accuracy). A total of 58 spectral indices are involved in the modeling. The Support vector machine-Recursive feature elimination (SVM-RFE) algorithm is used to screen the spectral index. This method uses the Support vector machine (SVM) algorithm to sort the feature variables and evaluate the importance of each feature variable. The variables with low importance are eliminated one by one according to the backward iteration, which can effectively remove the redundant features and improve the running speed of the model. A total of 36 machine learning models of different crops and different depths based on Random forest (RF), Extreme learning machine (ELM), Back-propagation neural network (BPNN) and other algorithms were constructed by taking the unfiltered full variable group and the filtered new variable group as the model input. The accuracy and inversion effect of each model were compared and evaluated. Finally, the best soil salinity inversion model and the best inversion depth under crop coverage were determined. The results show that the SVM-RFE variable selection algorithm can effectively improve the accuracy of each soil salinity inversion model. The R2 of the constructed model is higher than that of the model without variable screening, the RMSE is smaller, and the model training speed is faster. From the inversion model, on the whole, the inversion effect of RF model is better than that of ELM model and BPNN model, and the inversion effect can best reflect the real soil salt value. In 0-15 cm and 30-50 cm soil layers under crop cover, RF model had the best inversion effect, Rp2 of alfalfa field was 0.71 and 0.58 respectively, RMSEp was 0.026 and 0.033 respectively, Rp2 of corn field was 0.67 and 0.64 respectively, RMSEp was 0.111 and 0.094 respectively. In 15-30 cm layer, ELM model had the best inversion effect, Rp2 of alfalfa field was 0.58, RMSEp was 0.039, Rp2 of corn field was 0.68, RMSEp was 0.059. From the perspective of inversion depth, the inversion effect of 0-15 cm and 30-50 cm is better than that of 15-30 cm for alfalfa-covered soil, and the inversion effect of 0-15 cm and 15-30 cm is better than that of 30-50 cm for corn-covered soil. Comprehensive analysis, 0-15 cm is the best inversion depth for soil salt content under crop cover, the average R2 of the validation set is 0.65, and the RMSE is 0.084. The results of this study can provide a scientific basis for the rapid inversion of salt in different soil depth layers under crop mulching, and provide a reference for the management of saline-alkali land in the arid area of northwest China.