The assessment of above ground biomass density (ABGD) is essential for understanding the global carbon cycle and its impact on environmental dynamics. Despite advances in remote sensing technologies, the accurate estimation of biomass at fine spatial resolutions still presents challenges due to data gaps. Here, we propose a deep learning approach using Convolutional Neural Networks (CNNs) for global AGBD estimation at a 10-meter ground sampling distance. This approach is informed by near-real-time Sentinel-2 multispectral imagery and sparse GEDI LiDAR data. Our method adapts a CNN architecture initially created for canopy height mapping, is systematically assessed through various experiments considering geolocation, topographical, and climate data inputs. The best performing model achieves a mean absolute error of 35.9 Mg/ha and a root mean square error of 81.1 Mg/ha, showcasing competitive performance across continents against a global test set of over 300,000 samples. Notably, the inclusion of DEM elevations and geo-coordinates considerably improves AGBD predictions compared to the base model. The proposed method operates effectively without the need for extensive ground survey data, offering the potential for frequent updates to biomass density maps thanks to the revisit capability of Sentinel-2 satellites.