Reconstruction of Global Forest Aboveground Carbon Dynamics with Probabilistic Deep Learning

Understanding the long-term dynamics of forest aboveground carbon (AGC) is critical for constraining the terrestrial carbon cycle. However, accurately reconstructing historical AGC spatiotemporal patterns remains a challenge due to the complex, nonlinear relationships between vegetation proxies and biomass, as well as the stochastic uncertainties inherent in multi-source satellite observations.

In this study, we propose a probabilistic deep learning framework to reconstruct harmonized, high-resolution (0.25°) global forest AGC stocks and fluxes from 1988 to 2021. By integrating multi-source optical (e.g., NDVI, LAI) and microwave (e.g., VOD) remote sensing data, our approach utilizes Probabilistic Convolutional Neural Networks (CNNs) to simultaneously estimate AGC dynamics and quantify associated predictive uncertainties (decomposing aleatoric and epistemic components). This data-driven model effectively captures the nonlinear spatial dependencies and texture features that traditional empirical methods often miss.

Our reconstruction reveals significant decadal-scale regime shifts in the global forest carbon sink. While global forests remained a net sink of 6.2 PgC over the past three decades, we identify a pronounced transition in moist tropical and boreal forests, which have shifted from carbon sinks to sources since the early 2000s. Furthermore, our analysis uncovers an intensifying negative coupling between interannual tropical AGC fluxes and atmospheric CO₂ growth rates (r=-0.63 in the last decade), suggesting a growing complexity in the climate-carbon feedback. Spatially explicit partitioning in the Amazon further indicates a dynamical shift where AGC losses are increasingly driven by indirect climate stressors in previously "untouched" forests, rather than direct deforestation alone.

In conclusion, this study elucidates the state-dependent responses of global forests to changing disturbance regimes. The probabilistic framework provides a necessary basis for distinguishing genuine regime shifts, such as the structural decline of the tropical carbon sink, from observation noise, thereby enhancing our predictive understanding of terrestrial carbon resilience in a warming climate.