Digital mapping of soil salinization based on Sentinel-1 and Sentinel-2 data combined with machine learning algorithms

• Published in: Regional sustainability Vol. 2; no. 2; pp. 177 - 188
• Main Authors: Ma, Guolin, Ding, Jianli, Han, Lijng, Zhang, Zipeng, Ran, Si
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.04.2021; KeAi Communications Co. Ltd
• Subjects: Digital soil mapping; Ogan-Kuqa River Oasis; Salinization; Sentinel-1; Sentinel-2; XGBoost; Salinization; Ogan-Kuqa River Oasis; Digital soil mapping; XGBoost; Sentinel-1; Sentinel-2
• ISSN: 2666-660X; 2666-660X
• DOI: 10.1016/j.regsus.2021.06.001

Soil salinization is one of the most important causes of land degradation and desertification, especially in arid and semi-arid areas. The dynamic monitoring of soil salinization is of great significance to land management, agricultural activities, water quality, and sustainable development. The remote sensing images taken by the synthetic aperture radar (SAR) Sentinel-1 and the multispectral satellite Sentinel-2 with high resolution and short revisit period have the potential to monitor the spatial distribution of soil attribute information on a large area; however, there are limited studies on the combination of Sentinel-1 and Sentinel-2 for digital mapping of soil salinization. Therefore, in this study, we used topography indices derived from digital elevation model (DEM), SAR indices generated by Sentinel-1, and vegetation indices generated by Sentinel-2 to map soil salinization in the Ogan-Kuqa River Oasis located in the central and northern Tarim Basin in Xinjiang of China, and evaluated the potential of multi-source sensors to predict soil salinity. Using the soil electrical conductivity (EC) values of 70 ground sampling sites as the target variable and the optimal environmental factors as the predictive variable, we constructed three soil salinity inversion models based on classification and regression tree (CART), random forest (RF), and extreme gradient boosting (XGBoost). Then, we evaluated the prediction ability of different models through the five-fold cross validation. The prediction accuracy of XGBoost model is better than those of CART and RF, and soil salinity predicted by the three models has similar spatial distribution characteristics. Compared with the combination of topography indices and vegetation indices, the addition of SAR indices effectively improves the prediction accuracy of the model. In general, the method of soil salinity prediction based on multi-source sensor combination is better than that based on a single sensor. In addition, SAR indices, vegetation indices, and topography indices are all effective variables for soil salinity prediction. Weighted Difference Vegetation Index (WDVI) is designated as the most important variable in these variables, followed by DEM. The results showed that the high-resolution radar Sentinel-1 and multispectral Sentinel-2 have the potential to develop soil salinity prediction model.