Abstract
Water salinity characterizes the physicochemical properties of natural water, serving as an essential parameter for assessing lake water quality. However, the efficiency of remote sensing inversion of water salinity is limited as salinity is a non-optically active parameter, leading to the lack of a pixel-scale lake salinity dataset. Conventional function models based on salinity tracers or single lakes have low regional applicability, while machine learning algorithms can effectively capture the nonlinear relationship between radiance and salinity, providing large-scale inversion opportunities. Our study constructed an extreme gradient boosting (XGB) salinity model, which was used to generate the Inner Mongolia lake salinity (IMSAL) dataset with Sentinel-2 remote sensing reflectance. The IMSAL dataset contains 928 raster scenes with 10-meter spatial resolution for eight lakes from 2016 to 2024. Cross-validation and independent validation with measured and published literature-recorded salinities confirmed the good consistency and reliability. This dataset provides invaluable information on spatial patterns and long-term variations in lake salinity useful to prevent lake salinization and facilitate the lake management for sustainable ecosystem development.