Moso bamboo (Phyllostachys pubescens) is an introduced, invasive species of bamboo widespread across central and southwest Japan, where it can substantially impact carbon (C) cycling and forest management. Forest abandonment has led to the accelerated spread of Moso bamboo into adjacent ecosystems, where its rapid growth and high CO2 uptake create both ecological disruption and potentially a significant carbon sink.

We applied a Random Forest (RF) machine-learning model to estimate annual soil respiration (SR) from a Moso bamboo forest at Odake, in Sakura City, Chiba Prefecture. Monthly CO2 fluxes were measured using static chambers, alongside soil temperature and moisture, air temperature, humidity, and weather data from three sites in Chiba and Tochigi Prefectures.
The RF model was trained on this dataset, key predictor variables were identified, and model performance was evaluated using 10-fold cross-validation to ensure model stability and generalisability. Predicted SR closely matched the measured values, with no significant difference between simulated and observed CO2 fluxes.

The trained and validated model was then used to generate daily SR estimates for one year at the Odake site under three management scenarios: Clear-cut, Managed by thinning, and Abandoned. Annual SR totals were 6.38, 6.21 and 6.12 kg CO2 m-2 yr-1 respectively, with clear cut and thinned sites having significantly higher SR in the summer months.

These results demonstrate the utility of machine-learning approaches for evaluating management-driven variations in soil carbon fluxes in Moso bamboo forests, paving the way to addressing important questions for the long-term management of Japan’s forests.