Estimation of Salt Content of Saline Soil in Arid Areas Based on GF-5 Hyperspectral Image

There are few spaceborne hyperspectral sensors, and the estimation of soil salt content based on hyperspectral images is still under exploration. GF-5 is the satellite with the highest spectral resolution in China. This paper aims to study the feasibility of estimating salt content of saline soil in arid areas on a large area using GF-5 hyperspectral image. In this paper, 198 soil samples were collected from the experimental field at Yanqi, Xinjiang. Firstly, the soil salt contents (total salt content, Na+, Ca2+, SO42 and Cl) were determined, and the spectra of the soil samples were measured with an ASD Fieldspec3 field spectrometer in the laboratory. Then, the laboratory soil spectra were subjected to SG (Savitzky-Golay) smoothing pretreatment, and the competitive adaptive reweighted sampling method was used to select the characteristic bands of soil salt. Partial least squares, ridge regression and support vector machine established the regression model of soil salt content. It is found that the soil salt retrieval model established by laboratory spectra has high accuracy. The determination coefficients of the correction set and prediction set of the five soil salt retrieval models are greater than 0. 97 and 0. 90 respectively. Next, the GF-5 hyperspectral image data at the same time as soil sampling are obtained and preprocessed. The spectra of 198 soil samples were extracted from the image based on the location of the sampling points. Soil salt retrieval models based on GF-5 hyperspectral image spectra were established using the same retrieval method of laboratory spectra. The best prediction set determination coefficients of the five soil salt (total salt content, Na+, Ca2+, SO42, and Cl) retrieval models were 0. 76, 0. 66, 0. 76, 0. 63 and 0. 77 respectively. Finally, according to the retrieval results of soil salt based on the GF-5 image spectra, the characteristic band combination and modeling method with the best accuracy were selected estimate soil salt content in the whole study area. The estimation results have been divided according to the salinization grade. The saline soil in the study area accounts for 76%, and the land can not be cultivated. Non saline soil accounts for 16%, and crops can be planted. The distribution area of weak, medium and strong saline soil is small, accounting for 8% in total. The spatial distribution trend of the five soil salt estimation maps is consistent with the total salt content interpolation map. This paper shows that the results of estimating soil salt content in this study are based on GF-5 hyperspectral image are highly reliable.