Abstract: Soil salinization is one of key drivers for the degradation of soil quality and yield in arable land. To accurately and quickly evaluate soil salt content in Yinchuan Plain, field and indoor hyperspectral data were processed with first order differential (FDR) transformation, then the feature bands were identified by stepwise regression (SR). Partial least squares regression (PLSR) and support vector machines (SVM) were used to build models, which were verified to figure out the optimal hyperspectral type for the study area. Moreover, segmented and global corrections were performed to process poor hyperspectral, aiming to improve the accuracy of soil salt content inversion. The results showed that the accuracy of soil salt content inversion model based on field hyperspectral data was 58.9% higher than that of the indoor hyperspectral data. The accuracy of the inversion was improved through the segmented and global correction of the indoor hyperspectral. We found that the segmented correction is more accurate for the PLSR model (R_c²=0.790, R_p²=0.633, RPD=1.64) and the global correction is more accurate for the SVM model (R_c²=0.927, R_p²=0.947, RPD=3.87). The SVM models' inversion accuracy was higher than that of PLSR, with the field hyperspectral model fitted the best, followed by the indoor hyperspectral processed with the global correction and the indoor hyperspectral processed with the segmented correction, while the indoor hyperspectral the worst. Our results suggest that field hyperspectral data could contribute to the quantitative inversion of soil salt content in Yinchuan Plain. The corrected indoor hyperspectral could significantly enhance the inversion accuracy of soil salt content, which could guarantee food security and ecological quality development.

Key words: soil salt content, field hyperspectral, indoor hyperspectral, segmented correction, global correction