Quantifying mountain forest carbon stocks and elucidating their spatially heterogeneous driving mechanisms are both critical for terrestrial carbon management under the global carbon neutrality agenda. Conventional single-source remote sensing approaches can neither fully exploit multi-source data synergies nor adequately resolve spatial heterogeneity in complex terrains. This study develops an integrated framework combining multi-source remote sensing classification, InVEST-based carbon estimation, and multiscale geographically weighted regression (MGWR) and applies it to Zixi County, a subtropical mountainous bamboo-abundant region in southeastern China. Sentinel-2 imagery, PlanetScope data, and DEM derivatives were fused with an optimized Random Forest classifier, achieving an overall accuracy of 0.8565 (Kappa = 0.7065). Carbon stocks were then estimated via the InVEST model. MGWR analysis (adjusted R2 = 0.930, AICc = 594.032) substantially outperformed the global OLS model (adjusted R2 = 0.795, AICc = 1717.450), confirming strong spatial non-stationarity across all drivers. Canopy density exhibited the strongest positive local effect (coefficient range: 0.343–0.768); slope position showed predominantly negative regulation with localized positive reversals (−0.778 to 0.270); elevation displayed a broad-scale positive gradient (0.133–0.140); and total vegetation cover exhibited bidirectional effects (−0.134 to 0.208) with pronounced east–west divergence. This framework not only provides a robust methodological reference for carbon stock assessment in complex mountain landscapes but also supports targeted forest management and carbon sequestration strategies through spatially explicit driver identification.