Soil salinization seriously restricts the circular development of farmland economic production, and it is of great significance to monitor the dynamic change of soil salinity efficiently and accurately for the improvement and utilization of saline-alkali land. To timely and effectively monitor saline content in four typical salinized farmland improvement demonstration areas on the south bank of the Yellow River in Inner Mongolia, for example, using Sentinel-2 multispectral remote sensing image, synchronous collecting the surface soil salt data, screening sensitive spectral index through correlation analysis, based on three simple machine learning models of PLSR, SR and RR and Transformer deep learning model, finally precision evaluation and optimization of the best salt inversion model was carried out. The results showed that the visible light, red edge, and nearred band reflectance values of soil reflectivity in the demonstration area were positively correlated with soil salt content. The reflectivity values of the short-wave infrared band were negatively correlated with soil salt content. Introducing spectral index can effectively improve the correlation between Sentinel-2 remote sensing images and the salt content of the surface soil in the demonstration area (|r|≥0.32). A comparison of different models found that the Transformer deep learning model outperformed the simple machine learning model, and the R2 and RMSE of the validation set were 0.546 and 2.687g/kg;the salt inversion results were consistent with the field results, which provided a reference for more accurate inversion and improvement of the salinization degree in the south bank of the Yellow River in Inner Mongolia.