This study aims to improve the monitoring and prediction ability of soil particulate organic carbon (POC) in coastal salt marsh wetlands in response to sea level rise, and provide a scientific basis for carbon pool management in coastal wetlands. This paper constructs an ensemble learning sensitivity model based on multi-source environmental data, and compares the prediction performance of machine learning methods such as Random Forest (RF) and XGBoost. The data included multi-dimensional variables such as elevation, flooding frequency, pore water salinity, sedimentation rate, vegetation index and soil particle size. The data were imputed with missing values, excluded with outliers and standardized. Combined with the Shapley additive explanatory value (SHAP), the contribution of each environmental factor to the sensitivity of POC and regional differences were analyzed. The results show that the POC of coastal low-lying salt marshes is the most sensitive to sea rise disturbance, and the response of central salt marshes and inland high salt marshes is weakened in turn, showing obvious spatial heterogeneity. In the typical salt marsh area, the predicted R² of the model can reach 0.88, RMSE and MAE are 0.015 and 0.012, respectively, indicating that the ensemble learning method has good performance in capturing the nonlinear coupling and spatial heterogeneity of environmental variables. SHAP analysis showed that flooding frequency, deposition rate and pore water salinity were the key drivers of POC change, and their direction and strength varied with regional conditions. This study provides quantitative methods and decision support for blue carbon management, ecological restoration and digital monitoring of coastal wetlands.
