Lake wetlands play essential roles in flood regulation, water purification, and biodiversity, and restoring degraded wetlands can enhance their carbon-sink capacity. However, knowledge regarding the spatial patterns, temporal dynamics, and key drivers of soil organic carbon density (SOCD) in restored wetlands is limited. This study integrated multi-source remote sensing and machine-learning approaches to map SOCD in a restored lake wetland and to assess its spatiotemporal evolution from 2000 to 2023. Results showed that soil depth, soil physicochemical properties, and near-infrared spectral features were the major drivers of SOCD variability. Among nine models tested, CatBoost achieved the highest accuracy under random five-fold cross-validation (R2 = 0.49 f 0.06; MAE = 6.78 f 0.58 MgC center dot ha-1; RMSE = 9.30 f 1.17 MgC center dot ha-1; RPIQ = 1.56 f 0.17), whereas the RF model performed most robustly under spatially grouped cross-validation(R2 = 0.41 f 0.16, MAE = 7.45 f 0.33 MgC center dot ha-1; RMSE = 9.76 f 0.64 MgC center dot ha-1; RPIQ = 1.48 f 0.24), demonstrating the advantage of ensemble learning in capturing spatial heterogeneity. Spatially, SOCD decreased from the upland to the lake-shore zones. Temporally, SOCD declined from 2004 to 2017 due to water-level fluctuations and aquaculture disturbance, with high-SOCD areas shrinking by 38.34 km2. Following the "Return-Aquaculture-to-Wetland" restoration project launched in 2017, SOCD recovered markedly, reaching a two-decade maximum in 2022 (high-SOCD area: 125.54 km2; regional mean: 47.60 MgC center dot ha-1). These findings provide a scientific basis for wetland carbon-stock assessments and ecological-restoration monitoring and demonstrate the potential of combining multi-source remote sensing and machine learning for large-scale SOCD mapping.