Remote Sensing Monitoring Research Of Soil Salinization Based On Support Vector Machine Regression

Soil salinization is an important reason that affects the ecological environment of the region and leads to the decrease of grain production,which restricts the sustainable development of local agricultural production and national economy.Pingluo county belongs to the arid and semi-arid region in the northwest of China,in which the soil salinization is a typical and prominent problem.The estimation and inversion of soil salinity information content in this region can provide reference for other studies on soil salinization.In this study,the center part of pingluo county was taken as the research area.The soil salt content was estimated using the hyperspectral data measured near the ground and the data of gaofen-2 remote sensing image.Through wavelet transform denoising and numerical transformation of hyperspectral measured data,the reflectance data of the band from 400nm to 1800nm were used as the characteristic values to analyze the soil reflectance spectrum and the spectral characteristics after transformation.Feature bands were selected to establish a support vector machine regression(SVR)model.Combining the measured hyperspectral data to simulate the broadband reflectance and obtain the spectral index,a soil salinization inversion model based on SVR model was established.The main conclusions are as follows:(1)The spectral curves of soils with different salt content are similar,and the reflectivity increases gradually with the increase of wavelength in the band range of 400nm-1800nm.With the increase of total salt content(that is,the degree of salinization),the spectral reflectance of soil increases gradually in each band.(2)Wavelet analysis can effectively denoise spectral curves.By comparison,when the wavelet base is selected db3,the decomposition scale is selected 3 layers,and the threshold is selected H,the denoising effect of the salinized hyperspectral data is optimal,and the subsequent hyperspectral transformation analysis can be carried out.(3)The inversion estimation results of SVR model of soil salinity based on the measured hyperspectral reflectance data of soil are good.The correlation between the spectral data and the total soil salinity is obviously improved after the original measured hyperspectral reflectance data are transformed by reciprocal logarithmic differential,and the correlation coefficient reached-0.55-0.55.The sensitive bands of the original hyperspectral data are concentrated in 450nm-550nm.After the data transformation,the correlation between the hyperspectral data and the soil salt content was significantly enhanced,and the correlation was larger in the range of 800nm-1000nm.The prediction effect of the SVR model based on the measured hyperspectral data is good,and the correlation is very significant at the level of 0.01.(4)The correlation between vegetation index and salinity index of the original multispectral gf-2 satellite remote sensing image and the soil salinity was calculated and the sensitivity of each index was analyzed.Then the reflectance data of four multi-spectral bands(B,G,R,NIR)were simulated by the hyperspectral reflectance data of soil and the response function of band of gf-2 remote sensing image,and the vegetation index and salinity index based on the simulated band reflectance data,which were calculated the correlation with soil salt content.The sensitive band and sensitive index were obtained.The SVR model was used for modeling and estimation.The model results were further analyzed and tested,and R2=0.685,RMSE=0.949,p=1.08×10-7 were obtained,and the inversion accuracy is 75%.