Remote Sensing Inversion Model of Cultivated Land Salinity Based on Attention Mechanism and Lightweight CNN: Construction, Validation, and Multi-Model Comparative Analysis

Soil salinization threatens agriculture in arid regions, and remote sensing retrieval still faces challenges of unclear mechanisms and poor generalization. Based on Sentinel-2 data, this study compares the retrieval performance of various machine learning and deep learning models for farmland soil salt content, and introduces an attention mechanism for optimization. The main conclusions are as follows: (1) Random Forest achieved the highest accuracy among classical machine learning models (R2 = 0.334), while CNN performed better among deep learning models (R2 = 0.70), making it suitable for modeling scenarios with a single dominant salt type, well-defined spatial structures, and samples covering major environmental gradients. (2) In the problem of multispectral salinity retrieval, the complementarity of errors among base models was poor, preventing the ensemble model from fully leveraging its advantages. (3) Specific indices derived from near-infrared, red-edge, and blue–green bands performed well for sulfate-type salinity retrieval. (4) In the seed maize production area of Gansu, approximately 75.47% of farmland is non-saline, with severely saline land accounting for 1.42%; over the past decade, 74.84% of the area experienced a decrease in salt content, among which areas with a significant decline (accounting for 9.31%) corresponded consistently with regions where continuous engineering salt removal and microbial fertilizer management had been implemented for ten years. This demonstrates that, under conditions of limited ground samples, combining Sentinel-2 spectral information with moderate local spatial context can enhance the ability to detect salinity changes in relatively uniform irrigated areas.