Modeling CO2 fluxes in coastal wetlands of China using explainable sequence-based deep learning

Coastal wetlands, mainly encompassing mangroves and saltmarshes, are unignorable carbon sinks, significantly contributing to atmospheric CO 2 sequestration. However, accurate assessment of their carbon sink capacity is hindered by limited observational data and modeling advancements. Temporal sequence-based deep learning, including Long Short-Term Memory (LSTM) networks and Transformer models, offers great potential for improving coastal wetland CO 2 flux predictions by effectively accounting for environmental memory effects, thereby enhancing regional flux estimates. Utilizing eddy covariance flux measurements, this study developed LSTM and Transformer models to predict CO 2 fluxes in coastal wetlands, incorporating multiple satellite-derived land surface datasets, and examined the influence of memory length on predictive performance. Results indicate that both models show comparably satisfactory performance, excelling in predicting gross fluxes (mangroves: R 2 = 0.80 for gross primary production [GPP], 0.89 for ecosystem respiration [RE]; saltmarshes: R 2 = 0.91 for GPP, 0.92 for RE), though less effectively for net fluxes (mangroves: R 2 = 0.51 for net ecosystem exchange of CO 2 [NEE]; saltmarshes: R 2 = 0.69 for NEE). The LSTM model optimizes at intermediate memory lengths (e.g., 6 months), whereas the Transformer's optimal length varies unpredictably across the tested range. Feature importance analysis, employing an advanced gradient-based SHAP method, identified LAI, Ta, and Ta as key predictors for mangrove NEE, GPP, and RE, respectively, and LAI, FAPAR, and LAI for saltmarsh equivalents. Notably, RE exhibited longer temporal memory effects than GPP in both ecosystems, with NEE showing intermediate dependence. This research highlights the advancements of sequence-based deep learning in modeling coastal wetland CO 2 fluxes, offering promising avenues for improving large-scale flux estimations for coastal wetlands. • We investigate the feasibility of state-of-the-art temporal sequence-based models, i.e., LSTM and Transformer, for predicting CO 2 fluxes of coastal mangrove and saltmarsh wetlands. • Both models perform satisfactorily in predicting coastal wetland CO 2 fluxes, with better performance on the gross fluxes (GPP and RE) than on the net flux (NEE). • The memory length plays an importance role for both models and their optimal lengths for predicting CO 2 fluxes of mangrove and saltmarsh are different. • Explainable AI using a gradient-based SHAP approach that considered temporal dependence demonstrates feature importance rankings in predicting mangrove and saltmarsh CO 2 fluxes. • This study illustrated the advances of sequence-based deep learning in predicting coastal wetland CO 2 fluxes.