Uncertainty in the estimation of aboveground biomass in mangrove forests using multiple-scale sampling data

Abstract

Most previous works on aboveground biomass (AGB) estimation provide a single estimate of AGB rather than the probability distribution of the predicted values. However, the NGBoost algorithm provides a probabilistic regression and uncertainty estimation solution. In this study, we validate NGBoost for estimating AGB in mangrove forests in northeastern Vietnam. We use spectral bands and image indices extracted from WorldView-2 as independent variables and field data from eight plots as the basis for analysis. By applying a spatial scaling sampling strategy, we derived approximately 290 aggregated samples from the established plots, which served as the dependent variables in subsequent modeling. To augment the training dataset and capture a broader spatial context, window filters of varying sizes were applied, enabling the inclusion of adjacent pixels into the sampling matrix. NGBoost hyperparameters were optimized by the meta-heuristic Fox-inspired Optimization Algorithm using the Root Mean Square Error (RMSE) as the objective function. The trained model ended up at an RMSE of 1.8771, a Mean Absolute Error (MAE) of 1.2898, and an R 2 of 0.924. We interpreted the trained model and found that the Green Leaf Index is the most influential factor in AGB estimation, far more than the following factors. Finally, we used the trained model to estimate AGB and its probability distribution for the entire study area.

Bibliographic Information