Multi-output deep learning models for enhanced reliability of simultaneous tree above- and below-ground biomass predictions in tropical forests of Vietnam

The development and evaluation of new methods for the measurement, monitoring, and assessment of forest carbon biomass is necessary to quantify the ecosystem services provided by forests. To that end, multi-output deep learning (MODL) models were developed, cross-validated as alternative to the conventional weighted nonlinear seemingly unrelated regression (WNSUR) method for simultaneous prediction of tree aboveground biomass (AGB), tree belowground biomass (BGB), and total tree biomass (TB = AGB + BGB), while ensuring additivity, in two main tropical forest types – Dipterocarp Forest (DF) and Evergreen Broadleaf Forest (EBLF). A destructive sample of 175 trees was collected from 27 purposively selected plots in the Central Highlands ecoregion of Vietnam. The potential predictors of AGB, BGB and TB included four tree-level variables (diameter at breast height, DBH; tree height, H; wood density, WD; and crown area, CA), three stand-level variables (Forest type; basal area, BA; and stand density, N), and five environmental variables (mean annual rainfall, P; mean annual temperature, T; Soil type; Altitude; and Slope). The model utilizing DBH, CA, H, WD, BA, Altitude, P, and Forest type as predictors performed the best among the MODL models developed in this study. Compared to WNSUR models that used the same set of predictors and the dataset from the same forest types of DF or EBLF, the MODL models reduced the mean absolute percent error of tree AGB, BGB, and TB by up to 24.7 %, 96.5 %, and 9.4 %, respectively. The results suggest that the MODL algorithm can be applied on a diverse spatial scale, covering gradients of forest stand characteristics, climate conditions, soil properties, and topography, as it can incorporate complex numerical and categorical variables into the models without requiring a priori functions.