Using satellite estimates of aboveground biomass to assess carbon stocks in a mixed-management, semi-deciduous tropical forest in the Yucatan Peninsula

Information on the spatial distribution of forest aboveground biomass (AGB) and its uncertainty is important to evaluate management and conservation policies in tropical forests. However, the scarcity of field data and robust protocols to propagate uncertainty prevent a robust estimation through remote sensing. We upscaled AGB from field data to LiDAR, and to landscape scale using Sentinel-2 and ALOS-PALSAR through machine learning, propagated uncertainty using a Monte Carlo framework and explored the relative contributions of each sensor. Sentinel-2 outperformed ALOS-PALSAR (R-2 = 0.66, vs 0.50), however, the combination provided the best fit (R-2 = 0.70). The combined model explained 49% of the variation comparing against plots within the calibration area, and 17% outside, however, 94% of observations outside calibration area fell within the 95% confidence intervals. Finally, we partitioned the distribution of AGB in different management and conservation categories for evaluating the potential of different strategies for conserving carbon stock.

Automated summary

Machine learning was successfully used to upscale forest aboveground biomass from field data to remote sensing data, while uncertainty was effectively propagated and the relative contributions of each sensor were explored. Sentinel-2 outperformed ALOS-PALSAR in the model performance, but the combination of both sensors provided the best fit.