Comparing the allometric model to machine learning algorithms for aboveground biomass estimation in tropical forests

Estimating aboveground biomass (AGB) pools is critical for understanding the structural characteristics of forest ecosystems, as the quantitative analysis of biomass contributes to a better understanding of carbon sequestration and in studying the productivity within forest ecosystems. A more precise estimation of AGB is necessary for the estimation of the amount of carbon stored in an ecosystem that is needed for the valuation of the ecosystem services. Though regression models have been used previously for both tree-level and stand-level biomass estimations, supervised machine learning models have rarely been put to test in field-based estimations of AGB. This study was done in the tropical forests of Northeast India (two reserve forests in Tripura) to compare the allometric performance of weighted nonlinear regression model (NLM) to random forest regression (RFR), extreme gradient boost regression (XGB), and support vector machine regression (SVR) in estimating tree-level AGB. Due to their ease of measurement in field conditions and ensuring practical use, tree diameter at breast height (DBH) and tree height (H) were used as the independent variables to train the models. 506 trees representing 21 local tree species were inventoried. Then, the four models (NLM, RFR, XGB, and SVM) were trained separately on the data and compared to find an optimal method to estimate AGB using modeling efficiency of fit (MEfit) and Root Mean Squared Error (RMSE). The models were then validated using the predicted residual error sum of squares (PRESS) statistics. The best fit model in our study was using RFR, with a high MEfit (0.95–0.99), low RMSE (63.10–132.39 kg), and outperforming all other methods in the PRESS statistics, with the highest modeling efficiency of prediction (0.94–0.99) and lowest mean of absolute PRESS residuals (34.99–78.23). The results from Northeast India presented here provide preliminary evidence that RFR could be a reliable method for determining the field-level carbon stock and understanding the biomass dynamics for evergreen and moist deciduous tropical forests. The model accuracy may be constrained by a limited number of variables chosen to suit practical use in the field, and incorporating additional tree-level or stand-level variables such as crown diameter and stand basal area could enhance the robustness and generalizability of AGB estimation.