Abstract: [Purposes] To explore the application of remote sensing indices at different time sequences in soil salinity prediction and address the issues of period specificity and weak generalization ability of single-phase remote sensing data in soil salinity prediction. [Methods] The research is based on the Landsat 8/9 monthly time-series images of 2023,constructing six groups of different types of remote sensing index combinations,and generating single-phase combinations of May and September as comparisons.After screening redundant sample points using the leave-one-out method,a random forest model for predicting soil salinity is constructed.Meanwhile,the SHAP method is combined to analyze feature contributions,and principal component analysis is utilized to explore the influence of redundant sample points. [Findings] The time-series model outperforms the single-phase model.Through the feature coupling of vegetation,water and salt indices,combination 5 performs best in the time-series scenario.Time-series data integrate the full-cycle information of vegetation growth and water transfer,effectively overcoming the limitations of single-phase data.The multi-dimensional balanced feature combination can suppress the interference of redundant samples and improve the stability of the model.Different types of remote sensing indices have different responses and predictive contributions to salt content,and are closely related to the type of index. [Conclusions] The technical solution combining random forest with multi-dimensional time-series index feature coupling and redundant sample control is effective in soil salinity prediction and can provide reliable technical support for regional soil salinization monitoring.

Key words: time-series remote sensing, random forest, soil salinity prediction, feature coupling, redundancy control, SHAP