Using Multi-source and Multi-index Method to Assess Soil Salinity in Hetao Irrigation District in Inner Mongolia

Abstract

【Objective】 This study is to propose a method to improve the accuracy of remote sensing-based soil salinity inversion in the Urartian irrigation area within the Hetao irrigation region. 【Method】 To address the challenge of low accuracy associated with using single data source, indices and algorithms for soil salinity inversion, we propose to use spectral indices and polarimetric-combination indices as the variables. Partial least squares regression (PLSR), adaptive boosting (AdaBoost), and random forest regression (RF) models were used to construct soil salinity inversion model, and the optimal inversion model was identified through comprehensive evaluation based on the spatiotemporal variation of soil salinity measured from the Urartian irrigation from 2019 to October 2021. 【Result】 The PLSR and AdaBoost models worked better using spectral indices than using polarimetric-combination indices, while RF model was superior in using the polarimetric-combination indices than using the spectral indices. The PLSR model was most accurate for predicting soil salinity at the depth of 10 cm, with the coefficient of determination being 0.70. The AdaBoost model was the best for predicting soil salinity at the depth of 2cm, with the coefficient of determination being 0.74, while the RF model worked best when using the polarimetric combination indices to predict soil salinity at the depth of 2 cm, with the coefficient of determination being 0.64. Soil with high salinity was predominantly located in the southeast, while soil in the northwest and the center were only slightly salinized. 【Conclusion】 Combing the improved spectral index and the AdaBoost algorithm is more accurate in using the inversion model to predict the salinity of the surface soil in Urartian irrigation area in the Hetao irrigation district.

Keywords

• sentinel remote sensing data
• soil salinity
• spectral index
• polarization combination index
• machine learning