Soil salinization represents a primary manifestation of land degradation and presents a significant threat to sustainable agricultural development.  Remote sensing-based methodologies currently constitute the preferred approach for salinization monitoring.  Environmental factors’ spatial heterogeneity substantially constrains the modeling process in accurately capturing the soil salt content (SSC)-modeling factor relationship, thereby affecting monitoring accuracy.  This study proposes a classification modeling framework based on dominant salinization factors, establishing distinct remote sensing inversion models through categorization of soil texture and surface drainage conditions.  Results indicate that classification modeling substantially improves the capture of SSC-modeling factor relationships.  The efficacy of identical modeling indicators and methods varies significantly across different classification scenarios.  Among the three modeling approaches, random forest demonstrates superior overall robustness.  Of the three variable selection methods, light gradient boosting machine (LightGBM) shows the strongest compatibility with the modeling approaches.  The classification strategy significantly enhances model accuracy: compared to non-classified modeling (R²_V=0.62), the testing set R² increases by up to 24% (R²_V=0.77).  Models under poor surface drainage category demonstrate optimal performance, with coupled models achieving R²_C=0.82 (training set) and R²_V=0.77 (testing set).  This research provides valuable insights for remote sensing monitoring of soil salinization in precision agriculture contexts.

machine learning  /  composting  /  adaptive boosting  /  nitrogen loss  /  feature selection