Abstract:Soil salinization is one of the important factors affecting agricultural sustainable development, to get the accurate and timely soil salinity content, and realize precision monitoring of salinization, taking covered fields in the territory of Wuyuan County, Bayinnaoer City in Inner Mongolia Autonomous Region as the research object, exploring UAV multispectral remote sensing platform combined with machine learning model to estimate the feasibility of different depths soil salinity. Firstly, UAV equipped with five-band multi-spectral camera was used to acquire high spatio-temporal resolution remote sensing image data, and soil salinity data at different depths of the ground were collected synchronously. Pearson correlation coefficient method (PCC), extreme gradient boosting (XGBoost) and gray correlation analysis (GRA) were used to optimize the spectral index. Then decision tree (DT), back propagation neural network (BPNN), support vector machine (SVM) and random forest (RF) machine learning methods were used to establish inversion models of soil salinity in farmland with different depths under vegetation coverage. The results showed that scheme 3 (XGBoost-GRA) variable optimization method can effectively screen out the sensitive spectral index, and the accuracy of the optimized spectral index based on this method was higher than that of the inversion model constructed by using only PCC or XGBoost method. By comparing the performance of different modeling methods at different soil depths, it can be seen that the RF model of random forest had the best overall performance, and the other three inversion models had also achieved better prediction effect. The prediction effect of 0~20cm soil depth was the best among the three soil depths. Among them, the determination coefficient R2, root mean square error (RMSE) and ratio of performance to inter-quartile distance (RPIQ) of the model with the highest accuracy were 0.820, 0.044% and 2.273, respectively. Moreover, the spatial distribution map of 0~20cm soil salinity drawn based on the best inversion model could reflect the degree of soil salinization. The research result showed that the combination of feature variable optimization and machine learning model can better estimate the soil salt content based on the UAV remote sensing platform.
