A novel memory-based deep learning framework for reliable joint estimation of daily vegetation carbon fluxes

• A novel memory-based deep learning framework was proposed for joint estimation of daily GPP, RECO, and NEE. • The proposed model achieved robust daily carbon flux estimations with enhanced reliability. • The linear probes revealed the increased reliability of the proposed model through better correlation in predicting non-target variables, i.e. vegetation transpiration and canopy conductance. • The importance of antecedent climate drivers on daily GPP was evaluated. Reliable estimation of ecosystem carbon fluxes is essential to understanding terrestrial carbon-climate interactions under global change. However, most existing models overlooked the vegetation memory effect and failed to capture the complex nonlinear interactions between carbon fluxes and forcing variables across different time steps. These have been one of the most significant challenges for accurate and reliable carbon dynamic modelling. To overcome these challenges, a novel memory-based deep learning framework, DemLSTM (Dual Encoder Multi-task LSTM), is proposed for joint estimation of daily gross primary productivity (GPP), ecosystem respiration (RECO), and net ecosystem exchange (NEE). Observations from FLUXNET2015, AmeriFlux, and ICOS Warm Winter dataset were leveraged to comprehensively train and test the model. Results indicated that DemLSTM improved the estimation accuracy of all components contrasting with Random Forest, XGBoost and vanilla LSTM, with R 2 of 0.743, 0.625, and 0.471 for GPP, RECO, and NEE respectively in site-basedcross validation, with R 2 of 0.698, 0.594, and 0.497, respectively, in the generalization test. DemLSTM achieved robust daily flux estimations with enhanced reliability. The framework remarkedly reduced carbon closure bias (the absolute bias between RECO-GPP and NEE) across 9 PFTs, ensuring superior physical consistency in ecosystem carbon exchange modeling. Additionally, the linear probe revealed the strength of DemLSTM in predicting not-target variables, vegetation transpiration and canopy conductance, with higher correlation coefficients. This highlights the DemLSTM’s ability to capture the intrinsic carbon-water coupling process, thus enhancing estimation reliability. Furthermore, XAI-based analyses revealed the lagged contribution of meteorological drivers to daily GPP, antecedent radiation, VPD, and air temperature consistently impacts daily GPP; the effect of antecedent water availability depends on ecosystem type, with deep-rooted ecosystems showing weaker memory effect to soil moisture and precipitation. This study provides a valuable reference for developing carbon flux estimation models with enhanced accuracy and reliability.