Modelling Soil δ13C across the Tibetan Plateau Using Deep-Learning

Abstract

Soil carbon isotopes (δ13C) provide reliable insights for studying soil carbon turnover at a long-term scale. The Tibetan Plateau (TP), often referred as “the third pole of the earth”, is highly sensitive to global climate change, and exhibits an early warning signal of global warming. Although many studies detected soil δ13C variability at site scales, there is still a knowledge gap existing in the spatial pattern of soil δ13C across the TP. In this study, we compiled a database of 198 topsoil δ13C observations from published literatures and used a modified multi-layer perceptron (MLP) neural network algorithm to predict the spatial pattern of topsoil δ13C and β (indicating the decomposition rate of soil organic carbon (SOC), calculated as δ13C divided by logarithmically converted SOC) at 500m resolution. Results showed that MLP model effectively predicted topsoil δ13C with a model efficiency of 0.72 and a root mean square error of 1.16‰. Topsoil δ13C varied significantly across different ecosystem types (p < 0.001) with a mean δ13C of –25.89 ± 1.15‰ (mean ± standard deviation) for forests, –24.91 ± 1.03‰ for shrublands, –22.95 ± 1.44‰ for grasslands, and –18.88 ± 2.37‰ for deserts. Furthermore, there was an increasing trend of predicted δ13C from the southeastern to the northwestern TP, likely linked to vegetation type and climatic conditions. β values were low in the eastern TP and higher in the northern and northwestern TP, indicating faster SOC turnover rate in the east TP compared to the north and northwest. This study represents the first effort to develop a fine resolution product of topsoil δ13C and β across the TP, which could provide an independent, data-driven benchmark for biogeochemical cycling models to study SOC turnover and terrestrial carbon-climate feedback over the TP under climate change.