Drylands, which cover more than 40 % of the Earth's terrestrial surface, are highly vulnerable to water erosion because of their sparse vegetation and intense rainfall. Although the crucial role of vegetation restoration in reducing soil erosion is well established, the specific mechanisms through which vegetation influences hydrological connectivity and soil organic carbon (SOC) dynamics in arid ecosystems remain inadequately understood, which limits the effectiveness of current restoration strategies. In this study, controlled experiments, image-based hydrological connectivity quantification, and explainable machine learning models were integrated to determine the mechanisms through which (grass, shrubs, trees, and shrubs-grass) regulate functional connectivity parameters (total flow path length, TFL; total flow width, TFW; and average flow width, AFW) and their linkages to erosion and SOC loss. Grassland-covered plots presented the highest TFW (16.4 cm) and lowest TFL (760 cm) values, promoting sheet flow dominance and reducing sediment loss by 94.5 % compared with those in the bare plots. In contrast, the TFL values (1326 cm) of the shrub-covered plots were comparable to those of the bare plots (1360 cm), indicating intensified rill-driven erosion. Structural connectivity parameters failed to capture dynamic flow path changes induced by vegetation stems, whereas TFL and TFW effectively characterized the erosion potential. The extreme gradient boosting (XGBoost) model outperformed the random forest (RF) model in predicting sediment (NSE = 0.96) and SOC (NSE = 0.85) losses, with SHapley Additive exPlanations (SHAP) analysis identifying the flow velocity and hydrological connectivity parameters (TFW and AFW) as the dominant drivers of sediment and SOC losses, respectively. This study provides an innovative perspective for understanding and optimizing sediment transport processes on dryland hillslopes.