Integrating remote sensing data and Bayesian Hierarchical Spatial Models for improved estimation of above-ground biomass in the lower Mekong Basin

Abstract

Accurate estimation of aboveground biomass (AGB) is essential for monitoring forest carbon stocks and supporting climate mitigation strategies such as REDD+. This study evaluates the utility of Bayesian Hierarchical Spatial Models (BHSMs)—specifically Spatially Varying Intercept (SVI) and Spatially Varying Coefficient (SVC) models—for estimating AGB in Nghe An province, Vietnam. A total of 21 models were implemented across three spatial aggregation levels using field plot data from Vietnam’s National Forest Inventory (NFI) and three remote sensing (RS) products: University of Maryland GLAD Tree Canopy Height (TCH), ESA Climate Change Initiative (CCI) biomass, and NASA GEDI Level 4B (L4B) gridded AGB. The NFI data included 371 subplot measurements grouped into 111 clusters. Subplot-level AGB values were averaged by cluster to produce plot-level estimates, and high-AGB outliers ($>$300 Mg ha$^{-1}$) were removed to construct a third dataset. Square root-transformed AGB was used as the response variable in all models. Performance was evaluated using cross-validation RMSE, bias, and posterior predictive distribution metrics, including credible interval width and empirical coverage probability. SVI models consistently outperformed direct RS predictions, and even the intercept-only model yielded lower prediction error. The inclusion of RS covariates offered limited additional benefit, and posterior slope variances were consistently small with short effective ranges, suggesting minimal spatial non-stationarity in covariate effects. GEDI-based models exhibited the best overall performance, providing the lowest RMSE values and most reliable uncertainty estimates. These findings support the use of BHSMs for AGB estimation in data-limited tropical landscapes, where spatially structured intercept terms can model broad-scale variation in AGB not captured by RS covariates.