This study develops an innovative framework for estimating forest carbon stocks in Sanjiangyuan National Nature Reserve through synergistic inversion of multi-source remote sensing data. We pioneer the comprehensive fusion of Sentinel-2red-edge spectral features with Sentinel-1SAR datasets. Comparative analysis of machine learning algorithms-Random Forest (RF), Gradient Boosting Regression Trees (GBRT), and K-Nearest Neighbors (KNN)-was conducted, with Recursive Feature Elimination (RFE) optimizing feature selection. Key findings reveal: The GBRT-RFE integration achieved optimal performance (R²=0.84, RMSE=3.50Mg/hm²). Multi-source fusion significantly enhanced estimation accuracy across forest types (broadleaf: R²=0.88; coniferous: R²=0.92). Red-edge indices (e.g., EVI, CI) demonstrated highest feature importance, while SAR textures effectively compensated for optical data limitations in canopy penetration through structural information. The 30-m resolution carbon stock map generated by the optimized model showed values ranging from 6.06 to 39.73Mg/hm² (mean=20.85Mg/hm²), exhibiting distinct topographic and moisture gradients. Methodological breakthroughs include: A topography-adaptive feature selection framework for alpine regions, improving steep-slope (>25°) estimation accuracy by 18%~22%; First successful integration of Sentinel-2red-edge bands (B5-B7, B8A) with Sentinel-1dual-polarization (VV/VH) textures; Development of a regional-scale high-precision estimation model (validated against 126 field plots, RMSE < 15%). This research delivers robust technical support for carbon sink quantification, with direct applications in carbon trading and ecological compensation programs, advancing China's "Dual Carbon" goals (2030/2060).