Information on forest biomass is crucial for a better management and understanding of wildfires effects (pre- and post-fire analyses), as well as for a more accurate monitoring of carbon cycle dynamics. The current literature adopts two main approaches for the retrieving of biomass values: 1) predicting the biomass through allometric equations based on field measurements or Airborne Laser Scanning (ALS) metrics; and 2) using machine learning models based on spaceborne data sources (radar, optical, and laser), trained with field or ALS data. While approach 1) is usually spatially limited and has high costs; approach 2) mostly focuses on the Aboveground Biomass Density (AGBD) (i.e.: leaves, branches, trunk), disregarding the other components of the forest ecosystems as root biomass (BGBD), litter biomass (LBD) and shrubs biomass (SBD)) which leads to biomass underestimation.</p> <p>This article proposes a novel methodology capable of producing accurate predictions of the four previously mentioned biomass components, that integrates multi-source remote sensing observations acquired by synthetic aperture radars (ALOS-2 and Sentinel- 1), multispectral sensors (Sentinel-2), and ancillary data (on topography and land cover). The methodology resorts to an extreme gradient boosting regressor trained with field inventory data for the AGBD and BGBD, and ALS data for the LBD and SBD. It was demonstrated the importance of including all the proposed datasets and that reliable predictions were achieved. Namely, %RMSEs of 63.29%, 62.90%, 41.07%, 28.34% were achieved for the AGBD, BGBD, LBD, and SBD, respectively.