Quantifying and monitoring carbon sequestration is an important tool for developing global policies, helping with the emerging carbon credit market, and understanding climate change. Above-ground biomass (AGB) is a commonly used indicator that describes the amount of carbon that is stored above ground. The estimation of AGB can be done using direct or indirect methods. Direct methods involve the destruction of the trees unlike indirect methods. Allometric equations constitute an indirect method that is widely used and does not involve the destruction of trees to estimate AGB. However, it involves collecting data from forest inventories, which is time consuming and expensive. A cheaper and faster alternative provided by technological development is remote sensing estimation. In this alternative, data collected using the satellite (remote sensing data) are used together with field data, which can lead to more accurate AGB estimates.

In recent years, several machine learning models have been used to predict AGB, mainly the Random Forest (RF) algorithm. However, RF models present limitations in their performance when dealing with spatial data, as is often the case in AGB, by ignoring the spatial autocorrelation, leading to one of the limitations in their predictive performance. Thus, within this context, we use a hybrid method to estimate the target AGB in Sierra de la Culebra, Spain, that combines the RF algorithm and a Bayesian geostatistical model and uses as features variables from remote sensing data from the GEOSAT-2 satellite, such as reflectance bands, vegetation indices, and texture variables. In addition, in this work, we compare the predictive performance of this hybrid model with the predictive results obtained by using solely the RF model and the Bayesian geostatistical model.