Soil salinization, recognized as a significant form of land degradation, has emerged as a critical threat to global agricultural productivity. The accurate and automated identification, segmentation, and extraction of varying degrees of salinized soil at regional scales present a pressing scientific challenge. While deep learning has emerged as an innovative and efficient approach for remote sensing data processing, its applicability and potential in salinized soil monitoring remain largely unexplored. This study addressed the spectral heterogeneity problem in remote sensing monitoring of salinized soil by developing a sample dataset through feature variable fusion, integrating field measurements and multi-source remote sensing data. We conducted a comprehensive comparison of multiple deep learning network models (PSPNet, SegNet, and U-Net) across three distinct datasets (Dataset A, B, and C) to evaluate the impact of data composition on monitoring accuracy and the sensitivity of network models to data transformation. The results demonstrate that all models achieved over 0.8 accuracy in saline-alkali land extraction across datasets, with U-Net showing the lowest loss value and strongest predictive capability. Dataset C was proved to be the optimal training dataset. Validation using field data confirmed the effectiveness of deep learning models for soil salinization classification in the lower Shiyang River basin, achieving over 0.8 accuracy in distinguishing salinized from non-salinized soils. However, the overall classification accuracy reached 0.45, limited by fuzzy boundaries between severe, mild, and non-salinized soils. Compared with existing data products, our approach provides more accurate, higher-resolution results that better reflect actual field conditions. This study offers methodological insights and theoretical references for long-term, large-scale monitoring and trend prediction of land features with fuzzy boundaries using intelligent approaches.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号