Abstract:Selenium (Se) is part of the essential trace elements in the human body. People obtain selenium mainly through the consumption of agricultural products, and selenium in agricultural products mainly comes from the soil. Therefore, studying the content and distribution of selenium in the soil is very important to human health and crop production. However, the development of hypersensitive remote sensing technology has made it possible to estimate the content and distribution of selenium in soil in an efficient, low-cost and large-scale manner. However, the sensitivity of soil selenium content of spectra is weak, which seriously affects the accuracy of quantitative inversion of hypersecretion selenium content. In this study, 50 soil samples were systematically collected from the study area to analyse the selenium content of the soil samples, and the soil reflection spectral data were collected simultaneously; the Savitzky-Golay convolutional smoothing algorithm, multiple scattering corrections (MSC), first-order logarithmic differentiation (lg(R)-FD), standard normal variance correction (SNV), multiple scattering corrected first-order differentiation (MSC-FD) for raw spectra enhancement; application of the stable competitive adaptive benighted sampling (sCARS) algorithm combined with Pearson correlation analysis (PCC) for feature band selection; comparative analysis of partial least squares (PLS), support vector machine (SVM) and particle swarm optimisation support vector (PSO-SVM) models for the quantification of soil selenium content in hypersecretion The results showed that the sCARS algorithm was applied to the inversion. The results show that applying the sCARS algorithm to the spectrally enhanced regression model and combining Pearson correlation (PCC) to select the feature bands with greater sensitivity to soil selenium content can not only reduce the complexity of the hypersecretion prediction model for soil selenium content and effectively avoid the loss of a large amount of useful information, but also improve the inversion efficiency of the hypersecretion regression model; comparing the training and prediction sets of different regression models The comparison of the coefficient of determination R2 and root mean square error RMSE between the training and prediction sets of different regression models showed that the support vector (SVM) model had better prediction results and higher model stability than the partial least squares (PLSR) model, and the non-linear model was more suitable for the prediction of soil selenium content; the inversion accuracy and stability of the SVM model were improved by optimizing the kernel function and regularization parameters of the SVM through the particle swarm (PSO) algorithm; the MSC-PSO-SVM model (R2=0.53, RMSE=0.34) and MSC-FD model (R2=0.50, RMSE=0.04) had more outstanding prediction results. In summary: the hypersensitive quantitative inversion model of soil selenium content using sCARS combined with the PSO-SVM algorithm can provide a new way to estimate the soil selenium content in a large hypersensitive area.
Key words:Hyperspectral; Soil Se Content; sCARS; PSO-SVM