Bayesian geostatistical models for predicting above-ground biomass from remote sensing data

Monitoring and quantifying the amount of carbon stored in forests is crucial for shaping global policies, understanding climate change, and supporting emerging carbon credit markets. One key indicator that effectively describes carbon quantity is the Above-Ground Biomass (AGB), that has been traditionally estimated using allometric equations and data collected in forest inventory data. Recently, advances in technology have enabled the integration of remote sensing data in this process. In this study, Bayesian geostatistical models (using INLA and SPDE approaches) were applied to estimate AGB in two areas of the Atlantic Forest in northern Spain, Sierra de la Culebra and Montaña Palentina . Field data from the Cuarto Inventario Forestal Nacional (IFN4) of Spain were combined with satellite data from GEOSAT-2. Measurements such as texture, vegetation indices, and spectral variables were extracted from satellite reflectance bands and used as covariates in the models. The models incorporated a spatial random field to capture spatial autocorrelation, which was found to be significant in both areas. In Sierra de la Culebra , NDVI (Normalized Difference Vegetation Index), homogeneity, and Forest Type were statistically significant covariates in modelling AGB. In contrast, in Montaña Palentina , the Red and Blue reflectance bands were the main contributors to AGB estimation. Predictive results in Sierra de la Culebra were satisfactory (RMSE = 47.06 (t/ha); MSE = 2214.68 (t 2 /ha); MAE = 36.84 (t/ha); R 2 = 0.52; rRMSE = 39%), while in Montaña Palentina , predictive performance was even better (RMSE = 24.41 (t 0 . 75 /ha); MSE = 505.71 (t 1 . 5 /ha); MAE = 20.02 (t 0 . 75 /ha); R 2 = 0.5; rRMSE = 36%). • Estimation of AGB in two Areas in Northern Spain that belong to the Atlantic Forest. • The Bayesian geospatial model using INLA+SPDE was used to estimate the AGB. • Use of non-parametric smoothers to improve the predictive capacity of the model in estimating AGB.