Estimation model of desert soil organic matter content using hyperspectral data

Abstract: Little reports focus on estimating soils with organic matter content smaller than 2% using hyperspectral data. However, it is important for desert soils where most soil organic matter is low. Due to low organic matter content (<2%) in deserts soil, it is difficult to identify spectral features of soil organic matter and to determine sensitive bands. Therefore, this study aimed to establish a method for organic matter content of desert soils using Hyperspectral data. Thirty-two soil samples were collected from the Eastern Junggar Basin, China. The soil organic matter of these samples was determined. Meanwhile, the reflectance of samples was measured. The correlation between soil spectrum and its organic matter content was analyzed. Twenty-four of 32 samples was used for establishing hyperspectral models of estimating deserts soil organic matter content (SOMC) and the other 8 samples was for model verification. The results showed all the soil samples had the organic matter content <2%, in agreement with the study requirement here. Through comparing with response bands of soil organic matter with its content of 2% above, we found that spectral sensitive region 640-790 nm for deserts soil organic matter <2% was most sensitive. The correlations between reflectance spectra, inverse-log spectra (log (1/A)) and SOMC were low. Their correlations were improved after conducting the first-order differential and second-order differential transformation on soil spectra data. The correlation of partial spectrum bands by F test was significant (p<0.01). Thus it could be used to estimate deserts SOMC. The accuracy of linear regression model was rather low (R2<0.60) and not suitable to estimate deserts SOMC. Compared to the linear regression model, the R2 of second-order differential model and inverse-log second-order differential model based on multiple stepwise regressions was increased by 0.22 and 0.21, respectively, and their RMSE were decreased by 0.66 and 0.88, respectively. As such, the estimation accuracy of multiple stepwise regression models was higher than that from the linear regression model. Compared with the multiple stepwise regression models, the R2 of first-order differential model and second-order differential model based on partial least square regression was increased by 0.07 and 0.04 but the RMSE was reduced by 0.11. Therefore, the second-order differential model based on partial least square regression was the optimal model among the 12 estimation models established in this paper. For both multiple stepwise regressions and partial least square regression, the second-order differential estimation model was best in all the estimation models established. Therefore, the estimation of deserts SOMC using multiple stepwise regressions was feasible. The results here provide valuable information for hyperspectral remote sensing analysis on deserts soil organic matter, realize timeliness and accuracy of monitoring deserts soil organic matter, and helps with regional ecological environment restoration.