Comparing statistical and deep learning approaches for simultaneous prediction of stand-level above- and belowground...

Accurate and cost-effective prediction of aboveground biomass (AGB), belowground biomass (BGB), and the total (ABGB) at stand-level within tropical forests is crucial for effective forest ecological management and the provision of forest ecosystem services. Although there has been research on simultaneously fitting biomass equations for tree components, rather few studies focus on simultaneously predicting AGB and BGB at stand-level while maintaining additivity. We developed innovative Deep Learning Additive Models (DLAMs) for the simultaneous predictions of stand-level AGB, BGB, and ABGB integrating forest stand, ecological, and environmental factors as predictive covariates and compared them with conventional weighted nonlinear seemingly unrelated regression (WNSUR) and multivariate adaptive regression splines (MARS). Data for this study were collected from 121 plots distributed in two tropical forest types (dipterocarp and evergreen broadleaf) across five ecological regions of Vietnam, capturing three response variables (AGB, BGB, and ABGB), and 12 predictors. Factor analysis for mixed data was employed to identify the optimal covariates. Cross-validation results demonstrated that DLAMs substantially enhanced the reliability of simultaneous predictions of forest biomass components compared to the conventional WNSUR and MARS methods. The optimal DLAMs included seven predictive covariates (stand basal area (G), stand volume (V), mean annual temperature (T), elevation (EL), forest type (FT), average height (Hg), and soil group (SG)). They had mean absolute percent errors (MAPEs) of 6.3 %, 4.3 %, and 5.3 % for the simultaneous prediction of AGB, BGB, and ABGB, respectively. The MAPEs for the DLAMs approach were substantially lower than those for the WNSUR alternative by 2.9 %, 14.0 %, and 2.4 %, and lower than those for the MARS method by 4.3 %, 11.6 %, and 4.1 % for predicting AGB, BGB, and ABGB simultaneously, respectively. Conducting experiments in designing multi-input multi-output deep neural networks was essential for significantly improving the reliability of the simultaneous predictions from the DLAMs.