Research On UAV Multispectral Remote Sensing Model For Estimating Soil Salt Content

Soil salinization is the main problem restricting irrigated agriculture in arid and semi-arid regions.Timely and accurate acquisition the information of soil salt content(SSC)is of great significance to prevent and control soil salinization and construct the sustainable irrigated agriculture.UAV multispectral remote sensing can obtain the spectral information of soil and crop on a farm scale quickly and accurately,reflect its spectral characteristics,and provide a strong guarantee for estimation of SSC and monitoring its dynamic changes.Therefore,four typical plots with different salinization degrees in Shahaoqu irrigation area of Hetao Irrigation District were selected as the study areas in this paper.We collected sampling points at topsoil(0?10cm)before and after spring irrigation and points at different depths(0?10cm,10?20cm,20?40cm)in the month with crop covering.Simultaneously acquired UAV multispectral remote sensing images in these periods,extracted spectral reflectivity of six bands and calculated various spectral indices.Next,we analyzed the correlation between spectral variables(spectral reflectance of six bands and calculated spectral indices)and SSC in different periods,and used three spectral variable selection methods to screen out sensitive spectral variables.Then,we constructed different SSC estimation models with various modeling methods,evaluated the model accuracy by multifarious precision evaluation indicators,obtained the best SSC estimation models in each period after comparing,and then drawn the soil salinity maps of study area in different periods based on these best models.The main results obtained in this study are as follows:(1)The SSC estimation models before and after spring irrigation are established,and the soil salinity maps in each period based on the best estimation models are drawn.Variable selection methods such as Variable Importance in Projection(VIP),Competitive Adaptive Reweighted Sampling(CARS)and Genetic Algorithm(GA)can effectively screen out sensitive spectral variables,and the model established by GA has the best estimation effect,followed by CARS,and VIP has the worst effect.Among the estimation models before and after spring irrigation,the prediction accuracy of the BP neural network(BPNN)model with variable selection method is higher than that of the Multiple Linear regression(MLR)model,and compared with the CARS method and the VIP method,the BPNN model with GA selection method has higher prediction accuracy(R_P² are 0.78 and 0.80,respectively).From the soil salinity maps we can know that the salinization degrees of study area are mainly none and mild salinization,and the soil salt content before spring irrigation is higher than that after spring irrigation with the tendency that the higher is the soil salinity,the more obvious is the change of SSC due to the irrigation.(2)Estimation models of SSC at different depths in crop covered period are established with spectral reflectance.The correlation between soil salt content at different depths and the reflectance of each spectral band has passed the significance test,and it at the depth of10?20cm is higher than that of 0?10cm and 20?40cm.Comparing the estimation result of SSC,we can find that the accuracy of the models constructed by machine learning algorithms are higher than that of the MLR model.The model constructed by BPNN at the depth of 10?20cm has the best estimation effect(R_P² is 0.584),followed by ELM model,and SVM model the worst.The models constructed by extreme learning machine(ELM)at depths of 0?10cm and 20?40cm have the highest accuracy(R_P² are 0.424 and 0.278,respectively),followed by BPNN model,and the support vector machine(SVM)model is the worst.(3)Estimation models of SSC based on improved spectral index at different depths in the crop covered period are established,and the corresponding soil salinity maps at different depths are obtained.A new spectral indices combining the traditional spectral index and the Rededge band are established and written as‘improved spectral indices'.The Elastic-net algorithm(ENET)is used to screen spectral variables.It is easy to find that the selected spectral variables have basically passed the significance test,and the number of variables has been significantly reduced.Comparing the SSC estimation models with improved and traditional spectral variable combination,we can find that the accuracy of the former is higher than that of the latter.Among the best SSC inversion models(ENET-ELM)at different depths,the model at the soil depth of 10-20 cm has the highest estimation accuracy(R_P² and CC₂ are 0.783 and 0.875,respectively,and RMSE_P is 0.141%),followed by that of0?10 cm,and the estimation effect at 20?40 cm soil depth is the worst.We drawn the soil salinity maps by optimal estimation model with the improved spectral variable combination at different soil depths,and these maps can accurately elucidate the salinization degree in the test area,indicating that introducing the Rededge band to construct the new spectral index is practicable to estimate soil salt content.