Assessment of a probabilistic supervised machine learning method to estimate biomass expansion and conversion factors: a case study on cedar and pine trees

Abstract

Quantifying tree and forest biomass is crucial for formulating effective forest policy and management, given its role in human resource use and carbon storage. Forest biomass significantly contributes to environmental quality by absorbing carbon dioxide. Current research focuses on accurately determining biomass factors for various tree species, reflecting the emphasis on estimating and predicting tree biomass and carbon stocks. This study employed both standard nonlinear regression modeling (NLR) and Gaussian process regression (GPR), a machine learning method using artificial intelligence, to estimate and predict biomass expansion and conversion factors accurately. The case study included plantation forests and naturally occurring cedar and pine trees in Turkiye's Western Inner Anatolian Region and Goller Region (Northern Mediterranean Region). Nonlinear regression used the Levenberg-Marquardt optimization method, while GPR employed the radial basis function kernel. This dual approach allowed for assessing prediction uncertainties. The models constructed using GPR show superior performance compared to the NLR models for both biomass factors and species within the datasets used. According to the Furnival evaluation metric values, the accuracy of the NLR models was 1.05 to 1.34 times lower than that of the corresponding GPR models. The overall findings highlight the significant potential of GPR for accurately estimating and predicting biomass factors with high variances. This emphasizes its utility in modeling scenarios that require high flexibility, such as tree biomass prediction.