Article ID: D-24-00026
An accurate estimation of soil heterotrophic respiration (Rh) is crucial to separate autotrophic respiration (Ra) from soil respiration (Rs) and to quantify the soil carbon balance. In this study, spatiotemporal variation in Rh within an area of 0.09 ha was modeled by machine learning (ML) with Random Forest (RF) and Gradient Boosting Machine (GBM) algorithms, based on hourly Rh data measured with five automated chambers over two growing seasons in an immature deciduous forest in Hokkaido, Japan. Using the explanatory variables of soil temperature, soil moisture (water-filled pore space (WFPS) or volumetric soil water content), soil bulk density, soil carbon/nitrogen ratio (C/N), wind speed, and litter accumulation, ML models were much superior to conventional regression models using soil temperature and/or soil moisture and a multiple linear regression model using the same variables as in the ML models. In addition, the RF model performed better than the GBM model in all variable combinations. According to the RF model, soil temperature showed the highest importance in Rh variation among the variables, followed by bulk density. The RF model is promising for the gap-filling of missing Rh data and the accurate separation of Ra from Rs.
