Deep learning models for improved reliability of tree aboveground biomass prediction in the tropical evergreen broadleaf forests

Aboveground biomass (AGB) and carbon uptake of a forest are key ecological indicators for various technical and scientific applications and sustainable forest management. Deep Learning (DL) methods have been considered as alternative to regression techniques to increase the reliability of tree AGB prediction. The objectives were to develop DL models to predict AGB in the tropical evergreen broadleaf forests and compare DL models with traditional regression equations for their reliability in AGB prediction. A total of 968 individual trees were destructively sampled from fourteen 1-ha and twenty-six 0.2-ha plots distributed across five ecoregions of Viet Nam to get a dataset of tree predictors of diameter at breast height (DBH), tree height (H), wood density (WD) and the response variable of AGB along with forest stand factors of basal area (BA) and density (N); ecological and environmental variables such as ecoregion, slope, altitude, soil type, averaged annual temperature (T), averaged annual rainfall (P) and averaged dry season length. The DL models were developed using different combinations of variables selected by factor analysis for mixed data and compared with traditional regression equations by using cross-validation. Trees AGB in tropical rainforest predicted by DL models had significantly higher reliability than the conventional regression equations when both had the same input variables. Of the 16 developed DL models with 1 to 9 predictors, the model with 9 predictors (DBH, H, Ecoregion, Altitude, P, T, Soil type, N and WD) was the best DL model which reduced root mean square percent error (RMSPE) and mean absolute percent error (MAPE) by up to 7.7% and 6.1%, respectively, compared to traditional allometric equations. The DL models created in this study should be applied for measured tree data following factors of the forest stand, ecology, and environment in sampled plots to predict the tree AGB and total AGB, carbon on a large scale with variation in the value of these factors. Thus, we recommend that the DL models apply for the Measurement, Reporting, and Verification (MRV) system of the Reducing Emissions from Deforestation and forest Degradation (REDD+) program at a large regional level, national or territorial level scale.