Understanding and monitoring the quantity and spatial distribution of biomass on the planet is essential for understanding the global carbon cycle. The earth’s biomass holds thousands of gigatons of carbon, and forests hold most of the biomass on land. Despite this, estimating and mapping global biomass is a difficult task. Measuring biomass with field surveys is time and cost prohibitive, and other pertinent technology such as lidar is often spatially discontinuous and temporally inconsistent. The Sentinel-1 satellites carry Synthetic Aperture Radar (SAR) sensors and will image nearly the entire globe every 12 days. SAR has an ability to estimate biomass, but the data are often noisy. To improve the accuracy of biomass estimation from satellite based SAR, this thesis provides a deep fully convolutional neural network architecture that can examine the shapes and textures along with pixel intensity values in SAR imagery to estimate biomass. The network is trained using lidar derived biomass images. The network is able to achieve an RMSE of 11.84 Mg/Ha from 4 study areas across 3 Western U.S. States and showed similar accuracy of the lidar derived biomass images when compared to field surveys. This thesis suggests that convolutional neural networks can increase the utility of SAR imagery for biomass estimation.