Soil Salinity Estimation Based on Sentinel-1/2 Texture Features and Machine Learning

Soil salinization is a vital factor in global land degradation, seriously affecting sustainable agricultural development. Efficient monitoring of soil salinity using satellite remote sensing is critical for saline soil management. Currently, research on soil salinity extraction using satellite remote sensing primarily relies on the spectral information of remote sensing images, but insufficient consideration was given to the texture features of the imagery and the integration of spectral and texture information. To fully explore the effectiveness of texture features and the integration of texture features and spectral information in soil salinity estimation, experiments were conducted in Shahaoqu Irrigation Area, Inner Monglia, China from April to August 2019. To this end, the experiments utilized measured soil salinity data and the textural and spectral data from Sentinel-1/2. The effectiveness of Sentinel-1/2 texture features in soil salinity estimation was examined using out-of-bag score, and soil salinity inversion models were constructed based on the texture features, spectral information, and four machine learning models [random forest, Cubist, support vector machines (SVMs), and backpropagation (BP)]. The results indicated that Sentinel-1 texture features were more sensible to bare soil salinity (the top four most sensible texture features were HOM, ENT, COR, and CON), while Sentinel-2 texture features were more sensible to vegetated soil salinity (the top four were VAR, CON, HOM, and ENT). In addition, when combining texture features and spectral information, the models exhibited improved performance, and random forest had the best performance in both bare soil (
${R}^{{2}}$
= 0.688 and RMSE = 0.207) and vegetated soil (
${R}^{{2}}$
= 0.494 and RMSE = 0.304).