National-scale temporal estimation of South Korean Forest carbon stocks using a machine learning-based meta model

Abstract

The objective of the present study is to develop an integrated meta model, i.e. a surrogate model for process-based model for estimating forest carbon stocks in South Korea. Two meta models which cover for the process-based model were firstly developed—meta forest growth and meta forest biomass and dead organic matter carbon (FBDC) by adopting a multi-layer feedforward neural network (ML-FFNN) with a scaled conjugated gradient for training. Due to dependence of the forest carbon stocks on growth and climate variability, it was possible to interrelate the two meta models, and define a single integrated meta model. For the meta forest growth model, scientific uncertainties surrounding the driving mechanisms of tree growth in the process-based model increased model complexity, resulting in a relatively fair model performance (R2 = 0.776), compared to a near-perfect meta FBDC model performance (R2 = 0.997). The integrated meta model maintained an intermediate performance (R2 = 0.822). The integrated meta model did well to capture the spatial patterning of carbon stocks when compared to those of previous process-based models, although the former had a limited capacity to capture extreme highs associated with Quercus acutissima Carruth. As the integrated meta model was developed by using data including the impact of climate change, it is applicable to use for the forest management, for example, the implementation of Nationally Determined Contributions in the future.