Simulating global forest NEP by integrating TL-LUE model with deep learning

Accurately estimating forest net ecosystem productivity (NEP) is paramount for a deeper understanding of the terrestrial carbon cycle and ecosystem functioning. However, current models struggle to represent environmental stress responses and canopy heterogeneity simultaneously, limiting accuracy and transferability across forest types. To address these limitations, we propose TL-DenseNet, a hybrid framework that augments a two-leaf light use efficiency (TL-LUE) backbone with DenseNet (Densely connected convolutional networks) deep-learning subnetworks to estimate dynamic actual light use efficiency (LUE), integrating multi-source environmental drivers to characterize their complex nonlinear relationships with NEP. TL-DenseNet further decomposes NEP into sunlit leaf GPP (GPPsu), shaded leaf GPP (GPPsh), and ecosystem respiration (Re) and performs end-to-end NEP reconstruction. Results show that at the 8-day scale, the model reduces RMSE relative to the purely data-driven DenseNet by approximately 6.56%, 7.35%, 3.12%, and 12.40% for MF, ENF, DBF, and EBF, respectively; compared with the stepwise physical TL-Rh model, RMSE is reduced by about 26.95%, 30.59%, 30.67%, and 32.41% across those forest types. Replacing a static LUEmax with dynamically inferred LUE yielded an additional NEP RMSE reduction of ∼3.5%–6.6%. Across all flux tower sites, 89% exhibited high agreement between simulated and observed 8-day NEP seasonal variations (R2 > 0.6), demonstrating strong generalization performance. These results indicate that TL-DenseNet, by combining mechanistic process representation with data-driven optimization, substantially improves NEP estimation accuracy and generalizability while retaining physical interpretability, thereby providing a scalable methodological advance for global forest carbon-flux estimation.