Vegetation carbon storage estimation is an important part of natural resources monitoring. Remote sensing technology combined with ground sample can obtain the spatial continuous distribution of vegetation carbon reserves in the region, which makes up for the deficiency of the traditional artificial sampling survey estimation. However, most of the existing parameter and nonparameter remote sensing estimation models ignore the relationship between sample data variation and spatial autocorrelation. In this study, Landsat 8 OLI images were used as data sources to extract remote sensing variables, combined with the survey data of vegetation carbon storage in ShenZhen. Three metrics, including Akaike Information Criterion corrected (AICc), maximum spatial autocorrelation distance (MSAD), and Cross-Validation (CV), were used to determine the optimal bandwidth. Different geographically weighted regression (GWR) models inversing carbon reserve were constructed using Gaussian, Bi-square and Exponential as kernel function, respectively. Compared with the multiple linear regression (MLR), the optimal model was selected to make the spatial distribution map of vegetation carbon storage in Shenzhen. The results showed that the overall precision of the GWR models were better than that of the MLR model. The coefficient of determination (R²) of the GWR models were higher than that of MLR model, and the root mean square error (RMSE) and mean absolute error (MAE) were reduced significantly. The selection of bandwidth and kernel function can have significant influence on the estimation results of GWR model. Among the constructed GWR models, the GWR model with cross-validation determined bandwidth and exponential as the kernel was the best, with R² of 0.697 and RMSE of 10.437 Mg C/hm², which increased by 13.87% to 32.28% on model accuracy compared with other models. And the variable regression parameters of this model had significantly spatial non-stationarity, which could better reflect spatial heterogeneity. Combined with the optimal GWR model and the spatial distribution of vegetation types, the carbon storage of vegetation in Shenzhen was estimated, and the carbon storage values were between 1.63 and 60.95 Mg C/hm². The high and low carbon storage were mainly distributed in forest and grassland areas, which was basically consistent with the vegetation coverage in Shenzhen. Considering the spatial heterogeneity of variables, the GWR model based on bandwidth optimization can obtain more reasonable spatial distribution of vegetation carbon storage to some extent, which can provide a method and technical reference for remote sensing estimation of vegetation carbon storage in Shenzhen.