Modelling Biological Regions using Species and Environmental Data to Quantify National Scale Aboveground Forest Biomass

Abstract:

The REDD+ program (‘Reducing Emissions from Deforestation and Forest Degradation’) seeks to reduce carbon emissions from deforestation and degradation in developing countries. Accurate monitoring of forest carbon stocks has been a critical challenge to many REDD+ countries, which are often characterised by highly diverse forests, rapid land cover changes, rugged terrain, and strong climate gradients. Here we present a two-step modelling approach to estimate the national forest carbon stock of Nepal: (1) Coarse resolution modelling of biophysical constraints on the potential aboveground biomass (AGB) and (2) fine resolution modelling of the deviation between the predicted potential AGB and the observed AGB as a function of disturbance, land use and protective status. Nepal has an estimated forest area of 5.9 Mha, including some of the most carbon dense forests (500-1000 tha^-1 AGB). Using Nepal’s national forest inventory and topo-climatic parameters, we applied probability-based clustering to identify sites with similarity in species composition. The clustering thus accounted for the effect of the broad biophysical gradients on forest AGB. We assumed that inventory plots with maximum forest AGB within each cluster represent AGB potential. Then, the expected potential for all sites within each cluster was derived by weighting predicted affinity to a cluster and maximum forest AGB of plots in respective cluster. 20% of the plots (389) were set aside for validation. In the second step, we used a Random Forest to model the deviation based on plot level forest AGB and higher resolution predictors related to disturbance and protective status. Spatially explicit predictions of current forest AGB are then obtained from the predicted potential (step 1) and the predicted deviation (step 2). The two-step model performed reasonably well with RMSE of 91 tha⁻¹ . Compared to existing global estimates of Nepal’s forest biomass, our approach reduced RMSE by more than half. The proposed approach offers an objective methodology to characterize spatial variability and quantify forest carbon in highly heterogeneous environments. We expect that the future availability of additional observations of forest structure as well as higher resolution climate grids will further reduce uncertainties in national forest carbon accounting.