Regionally trained models for mapping aboveground biomass from Remote Sensing data fusion: a comparison of the capabilities of Machine Learning in 4 different biomes.

In recent years, multiple remote sensing technologies have been used to quantify aboveground biomass (AGB) at different scales, from regionally trained models to global maps. The former are capable of providing higher resolution and accurate estimations albeit for smaller regions. Tailoring model weights and inputs to a specific biome or region have proved to be effective to obtain better results. Global maps, on the other hand, provide an attempt to standardizing AGB mapping at slightly to moderately lower resolutions and tend to have differences between them. A similar standardization with the benefits of regional mapping, applied across biomes, has yet to be available. For that, the first step would be comprehensively comparing state-of-the-art regional and local studies. However, existing inconsistencies in the different models and inputs used, which are often region-specific, make it impracticable.

This study addresses the need for a comparison of a single methodology consistent across different biomes in order to understand the nuances that drive the estimation of AGB. We present a data fusion approach to mapping aboveground biomass at a 20m resolution using regionally trained, regression-enhanced Random Forests in 4 different biomes: semi-arid savannas in the Sahel region, dense tropical forests in Brazil, Mediterranean forests in coastal and pre-coastal areas of Catalonia, and temperate/boreal forests in Minnesota. GEDI L4A AGB data (25-m discrete footprints) is used to train a regression model in each study area. We derived predictors from Sentinel-1 SAR, Sentinel-2 multi-spectral and Digital Elevation Model (DEM) datasets, which are common for all locations. Additionally, auxiliary data such as proximity to coastlines, human-made structures or bio-climatic variables are used to enhance predictions in saturation-prone areas. Additional to the goodness of adjustment to the training data from GEDI, we carried out a thorough validation of the results using in-situ data from Forest Inventory plots gathered in all 4 study regions. This enables a comprehensive comparison of the capabilities of Machine Learning modelling to adapt to the particular characteristics of each ecosystem. An in-depth analysis is carried out to find the most important predictor variables in each biome, as well as to assess the accuracy that can be expected across a wide range of AGB values.