The quantification and monitoring of above-ground grass carbon stock (AGGCS) will inform emission reduction policies and aid in minimising the risks associated with future climate change. This study investigated the sensitivity of Synthetic Aperture Radar (SAR)-derived parameters to predict AGGCS in a savannah ecosystem in Kruger National Park, South Africa. Particularly, we investigated the capabilities of Sentinel-1 derived parameters, including backscatter coefficients, intensity ratios, normalised radar backscatter, arithmetic computations, and the XGBoost tree-based algorithm, to predict the AGGCS. We further tested if incorporating texture matrices (i.e. Gray Level Co-Occurrence Matrix) can enhance the predictive capability of the models. We found that the linear polarisation (i.e. VV) and the intensity ratio (i.e. VH/VV) achieved similar results (R2 = 0.38, RMSE% = 31%, MAE = 6.87) and (R2 = 0.37, RMSE = 31%, MAE = 8.80) respectively. The Radar Vegetation Index (RVI) performed marginally (1%) better (R2 = 0.39, RMSE = 30% and MAE = 6.77) compared to the other variables. Nevertheless, the incorporation texture matrix into the model enhanced prediction capability by approximately 20% (R2 = 0.60, RMSE% = 20%, MAE = 3.91). Furthermore, the most influential predictors for AGGCS estimation were RVI, VHcor and VVcor order of importance. These findings (R2 values of 0.35–0.39) suggest that SAR data alone does not fully capture the variability in above-ground grass carbon stock, particularly in the complexly configured savannah ecosystems. Nevertheless, the results further suggest that the prediction accuracy of SAR-based above-ground grass carbon stock models can be enhanced with the incorporation of texture matrices.