Soil salinization monitoring in the Werigan-Kuqa Oasis, China, based on a Three-Dimensional Feature Space Model with Machine Learning Algorithm.

Abstract

ABSTRACT Land surface temperature (LST) is an important indicator for monitoring soil salinization in arid and semiarid regions. However, traditional salinity index (SI)-vegetation index (VI) feature space, and other single feature space often ignore the LST information, which has largely constrained the soil salinity detecting quantitatively and effectively. In this study, random forest (RF) as one of the machine learning algorithms was utilized to downscale the LST retrieved from Landsat 8 thermal-infrared band from 100 m to 30 m. Then, we developed a new three-dimensional feature space model that combines downscaled LST, SI and normalized difference vegetation index (NDVI) to assess the soil salinity in the Werigan-Kuqa Oasis, a typical delta oasis in an arid region, China. The experiment results are further compared with 20 different feature spaces and spectral indices. The result shows that the proposed model produces higher accuracy than other method, which can provide a rapid and relatively accurate monitoring results of soil salinization in the study area.