Interpretable machine learning reveals environmental controls on carbon-water flux dynamics in a humid subtropical evergreen forest
- 1State Key Laboratory of Water Disaster Prevention, Hohai University, Nanjing, Jiangsu 210024, China; College of Hydrology and Water Resources, Hohai University, Nanjing, Jiangsu 210024, China; Yangtze Institute for Conservation and Development, Nanjing, Jiangsu 210024, China
Abstract
Carbon-water coupling in humid subtropical forests is shaped by monsoon variability, yet coordinated changes in carbon uptake, respiration, and water loss remain unclear. We analyzed daily eddy-covariance and meteorological observations across six growing seasons (2020–2025) at Xin'anjiang, Southeast China, including net ecosystem exchange (NEE), gross primary productivity (GPP), ecosystem respiration (R eco ), evapotranspiration (ET), and water-use efficiency (WUE). To distinguish predictive contributions, nonlinear model responses, and covariate-adjusted associations under correlated predictors, we combined nested leave-one-growing-season-out XGBoost with SHAP, ALE, and EconML and paired gap-filled/non-gap-filled analyses. The forest remained a net carbon sink, but indicators varied asynchronously. During DOY 160–190, clustered rainfall and high humidity coincided with lower radiation; reduced GPP and net carbon uptake, comparatively high ET, and the seasonal WUE minimum. Interannually, NEE reflected the balance between GPP and R eco , whereas WUE depended on relative changes in GPP and total ET. Out-of-year R 2 ranged from 0.66 to 0.82, and the schemes shared 70–90% of their Top-10 predictors. Current atmospheric-energy conditions dominated NEE, GPP, and ET predictions; R eco was associated with temperature and antecedent moisture, whereas WUE retained information from atmospheric pressure and soil moisture. Nonlinear responses were clearest for radiation, temperature, and atmospheric pressure, while pairwise interactions were generally weak and covariate-adjusted associations were consistent across schemes. Together, monsoon-related carbon-water variation reflected immediate atmospheric-energy variability conditioned by slower hydrological memory. Although site-specific quantitative results require multi-site evaluation, this multi-flux framework may inform carbon-water-coupling studies in other monsoon-affected forests.
Related
- Mapping and conservation analysis of upper montane evergreen broad‐leaved forest using machine learning and geodata
- Sentinel-1 SAR time series for forest phenology: multimodal deep learning with meteorological fusion across deciduous broadleaf and evergreen needleleaf forests
- Differentiable land model reveals global environmental controls on latent ecological functions
- Environmental controls on the detectability of hydrodynamic changes induced by offshore wind farm structures
- Energetic Controls on Biodegradation of Organic Matter from Diverse Environmental Sources.
BibTeX
@article{zhang2026interpretable,
title = {{Interpretable machine learning reveals environmental controls on carbon-water flux dynamics in a humid subtropical evergreen forest}},
author = {Pengfei Zhang and Ke Zhang and Jiefan Niu},
journal = {Ecological Indicators},
year = {2026},
month = oct,
doi = {10.1016/j.ecolind.2026.115693},
url = {https://doi.org/10.1016/j.ecolind.2026.115693},
note = {CC BY},
}
RIS
TY - JOUR
TI - Interpretable machine learning reveals environmental controls on carbon-water flux dynamics in a humid subtropical evergreen forest
AU - Pengfei Zhang
AU - Ke Zhang
AU - Jiefan Niu
PY - 2026
DA - 2026/10/30
JO - Ecological Indicators
DO - 10.1016/j.ecolind.2026.115693
UR - https://doi.org/10.1016/j.ecolind.2026.115693
AB - Carbon-water coupling in humid subtropical forests is shaped by monsoon variability, yet coordinated changes in carbon uptake, respiration, and water loss remain unclear. We analyzed daily eddy-covariance and meteorological observations across six growing seasons (2020–2025) at Xin'anjiang, Southeast China, including net ecosystem exchange (NEE), gross primary productivity (GPP), ecosystem respiration (R eco ), evapotranspiration (ET), and water-use efficiency (WUE). To distinguish predictive contributions, nonlinear model responses, and covariate-adjusted associations under correlated predictors, we combined nested leave-one-growing-season-out XGBoost with SHAP, ALE, and EconML and paired gap-filled/non-gap-filled analyses. The forest remained a net carbon sink, but indicators varied asynchronously. During DOY 160–190, clustered rainfall and high humidity coincided with lower radiation; reduced GPP and net carbon uptake, comparatively high ET, and the seasonal WUE minimum. Interannually, NEE reflected the balance between GPP and R eco , whereas WUE depended on relative changes in GPP and total ET. Out-of-year R 2 ranged from 0.66 to 0.82, and the schemes shared 70–90% of their Top-10 predictors. Current atmospheric-energy conditions dominated NEE, GPP, and ET predictions; R eco was associated with temperature and antecedent moisture, whereas WUE retained information from atmospheric pressure and soil moisture. Nonlinear responses were clearest for radiation, temperature, and atmospheric pressure, while pairwise interactions were generally weak and covariate-adjusted associations were consistent across schemes. Together, monsoon-related carbon-water variation reflected immediate atmospheric-energy variability conditioned by slower hydrological memory. Although site-specific quantitative results require multi-site evaluation, this multi-flux framework may inform carbon-water-coupling studies in other monsoon-affected forests.
SN - 1470-160X
SN - 1872-7034
ER -