Validation of GEDI above ground biomass product over Pinus sp. plantations in Argentina

In recent years, in addition to the traditional uses of forest plantations, there has been an interest in their capacity to capture and fix carbon in their structures, acting as greenhouse gas sinks. The Global Ecosystem Dynamics Investigation (GEDI) project features a light detection and ranging sensor aimed at quantifying the distribution of biomass and carbon in woody vegetation worldwide. Although the GEDI level 4A product (L4A) generates aboveground biomass density (AGBD) estimates on a global scale using semi-empirical relationships between the energy reflected at a given height (rh metric) and AGBD, its forest segmentation is coarse and presents limitations at a local scale, which can result in less accurate estimates. We evaluated the accuracy of GEDI L4A AGBD estimates and in situ AGBD data obtained from forest inventories. The analysis of the relationship between in situ AGBD and GEDI L4A product AGBD estimates shows that, although the correlation is generally good, some plots exhibit high variability in GEDI AGBD values despite having intermediate values of in situ AGBD [root mean squared error (RMSE) = 89.1 Mg/h]. Moreover, we found that these underestimations are associated with low values of rh50. In particular, these low values of rh50 (related to large AGBD RMSE errors) were not related to thinned/non-thinned forests but were related to understory conditions (less error for the “very dense” understories). Then, we explored landcover misassignment; as the Pinus sp. plantations correspond to implanted North American evergreen needleleaf forests (ENFs), we reconstructed the L4A product for this region and South America, using the corresponding expressions to estimate AGBD. The results show an improvement in the overall correlation and a reduction in errors, with RMSE reducing to 68.4 Mg/h for South American ENF and 64.2 Mg/h for North American ENF. Finally, we discuss ways to improve AGBD estimations (or flag wrong ones) using all available values of the rh curve.