Conference Agenda

In recent years an increase in development and publishing of large scale (pan-troical to global) remote sensing based aboveground biomass (AGB) maps is observed. These datasets differ in spatial resolution, input data and modelling approaches. Given these differences it is important to compare the datasets with a common global reference data and understand implications for different use cases.

Here we present and overview and preliminary results from an ongoing validation of selected pan-tropical and global biomass products using AGBref, “A global forest biomass reference dataset”, and pan-tropical Airborne Laser Scanning (ALS) data collected in Brazilian Amazon, Democratic Republic of Congo and in Borneo, Indonesia. To do this, we compiled 23 AGB maps, matched each mapped year with AGBref, harmonized the forest definition by applying a consistent forest mask, disaggregated to 0.1 degree pixel spacing using mean values and calculated statistical metrics (i.e., bias, RMSE) at different biomass intervals. Furthermore, we used ALS-based AGB estimates to assess how well the global maps represent local spatial patterns. For this, we used high resolution AGB maps (i.e., at least 100 m pixel spacing) and calculated a local heterogeneity ratio index, defined as the ratio of standard deviation (SD) from a map within a moving window to the SD from the ALS-AGB within the same moving window.

Preliminary results indicate that the maps exhibit systematic differences, either as a uniform bias (e.g. overestimation across all biomass bins) or as over- and under-estimation in the lower and upper biomass ranges. An important trend is observed is that these biases tend to decrease in more recently produced maps. Furthermore, product versioning (e.g., the first vs. the last version) leads to a lower biases at lower and upper biomass levels. Finally, using ALS-AGB as reference, we show how much spatial details are captured, highlighting that recent maps based on high-resolution satellite imagery (<30m) and novel modelling approaches (e.g. deep learning) do not always preserve fine-scale spatial details.