Mapping soil salinity in the Weigan River Basin using sentinel-2 imagery and machine learning

Abstract

【Background and Objective】Soil salinization in Weigan River Basin of Xinjiang poses a significant pressure on agricultural production and ecosystem functions. Timely and accurate monitoring of salt content in the top 0-30 cm soil layer is essential for improving management and mitigation of salt-affected soils. We propose an inversion model in this paper to assess spatiotemporal variation in soil salinity across the basin.【Method】Using Sentinel-2 multispectral imagery and synchronously collected soil salinity data, three machine learning models: decision tree (DT), gradient boosted decision tree (GBDT), and random forest (RF), were constructed and evaluated to calculate soil salinity. We compared the three models first and then selected the most accurate one to calculate the complex relationships between spectral indices and soil salinity. This relationship was used to analyze the spatiotemporal variation in salinization in the basin.【Result】Strong correlations between soil reflectance spectral indices: NDVI, SRSI, and SI, and soil salinity were found, with the correlation coefficients＞0.6. Among all models we compared, the RF-based model was most accurate, with its associated R2, MAE and RMSE being 0.613, 2.951 g/kg and 4.524 g/kg, respectively. Spatial analysis showed that non-saline soil areas in the basin have increased by 2 404.17 km2, while moderately and weakly saline soil areas have decreased by 2 856.21 km2 and 2 518.06 km2, respectively. Despite the decline in soil salinity areas, mildly and heavily saline areas had increased by 1 935.51 km2 and 1 034.59 km2, respectively, highlighting that managing and remediating soil salinization is still a work not yet done.【Conclusion】The RF model combined with Sentinel-2 spectral indices provides an effective and accurate method for monitoring soil salinization at large scales. It can help develop strategies to remediate and manage soil salinity in the studied region.