Identifying early warning signals for carbon sinks in China's terrestrial ecosystems using deep learning

Terrestrial ecosystems in China play a critical role in the global carbon budget, yet their stability under ongoing climate change remains uncertain. While statistical indicators (e.g. AR1, variance) can detect a generic loss of resilience, they fail to classify the specific nature of potential critical transitions. In this study, we integrated multi-source data and a deep learning framework to analyze the early warning signals of China’s terrestrial Net Ecosystem Productivity (NEP) from 2001 to 2020. Specifically, we employed a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model trained on 100,000 synthetic time series to classify potential bifurcation mechanisms (Fold, Hopf, Transcritical) inherent in carbon flux dynamics. Then we applied XGBoost models and SHAP to reveal the primary predictors behind different bifurcation types. Our results show that while China’s terrestrial ecosystems functioned primarily as a carbon sink with an increasing NEP trend, indicators like AR1 and variance have risen in the subtropical and tropical monsoon zones. The CNN-LSTM classifier identified transcritical bifurcation as the dominant regime shift mechanism, affecting 59.4% of the study area, suggesting a prevalence of smooth rather than abrupt transitions. Abrupt "tipping points" (Fold bifurcations) and oscillatory instabilities (Hopf bifurcations) were relatively rare. Attribution analysis revealed that vegetation structure (Leaf Area Index) and hydrothermal conditions (precipitation and relative humidity) are the primary correlative predictors of these regime shifts. These findings provide a spatially explicit risk assessment for carbon sink destabilization, indicating that while the risk of abrupt catastrophic collapse remains low, widespread gradual regime shifts strongly linked to climate-vegetation feedbacks are underway.