Estimating Soil Salinity Using Multiple Spectral Indexes and Machine Learning Algorithm in Songnen Plain, China

Abstract

Soil salinization is one of the most critical threats to land degradation in arid and semiarid regions. This study is based on machine learning using Landsat 8 operational land imager (OLI) imagery to estimate soil salinity in Da'an City. A total of 19 spectral indexes, including 15 salinity indexes, 3 vegetation indexes, and a brightness index, were calculated using the blue, green, red, and near-infrared bands of Landsat 8 OLI images. Four machine learning regression algorithms, namely Cubist, support vector regression, random forest regression, and extreme gradient boosting regression, were used on the basis of the 19 aforementioned indexes to estimate soil salinity. Results demonstrated that the Cubist model has the highest prediction accuracy (RMSE = 0.31 mS/cm). Thus, the spatial distribution of soil salinity based on the Cubist model best meets the expectations of the authors. Moreover, the canopy salinity index correlated the most with the measured electrical conductivity, demonstrating a correlation coefficient of −0.44. After using the random forest method for variable screening, the Cubist model based on nine spectral indexes still achieved satisfactory prediction accuracy with the RMSE of 0.34 mS/cm. Thus, the Cubist method is recommended for soil salinity monitoring in arid and semiarid areas.

Keywords

• Digital soil mapping
• machine learning
• remote sensing
• soil salinization