Abstract: Soil salinization is one of the important factors restricting agricultural production, food security, and sustainable development in arid and semi-arid areas of China. With Yinchuan Plain as the research area, we tried to improve the way for hyperspectral data inversion of soil salinity based on data of hyperspectral reflectance and measured soil salinity. The original reflectance was processed with fractional-order differential transformation. Multiple linear regression (MLR) and support vector machine (SVM) models of soil salinity (with a sample size of 133) were established by using a twodimensional spectral index to screen sensitive parameters. Differential transformation can effectively enhance the response of hyperspectral information to soil salinity, and the integer order differential of first order had the best effect, with a correlation coefficient of 0.486. The effect of fractional-order differential order 0.9 was the best, with a correlation coefficient of 0.461. The combination of characteristic bands screened by fractional-order differential and two-dimensional spectral index had a stronger correlation with soil salinity than that of one-dimensional bands. The accuracy of the SVM model was better than that of the MLR model. The accuracy of the two models with order 1.1 was the best, with the R_p², RMSE and RPD being 0.839, 0.96, and 2.46 for the SVM model, and 0.730, 1.32, and 1.79 for MLR model, respectively. The inverse distance weight method of the SVM interpolation can better predict soil salinity in Yinchuan Plain, providing scientific support for improving the accuracy of hyperspectral inversion of soil salinity and a basis for accurate diagnosis, improvement and utilization of saline-alkali  lands.

Key words: soil salinity, fractional-order differential, spectral index, hyperspectrum.