Exploring the Potential of Long Short‐Term Memory Networks for Predicting Net CO2Exchange Across Various Ecosystems With Multi‐Source Data

Upscaling flux tower measurements based on machine learning (ML) algorithms is an essential approach for large‐scale net ecosystem CO2exchange (NEE) estimation, but existing ML upscaling methods face some challenges, particularly in capturing NEE interannual variations (IAVs) that may relate to lagged effects. With the capacity to characterize temporal memory effects, the Long Short‐Term Memory (LSTM) networks are expected to help solve this problem. Here we explored the potential of LSTM for predicting NEE across various ecosystems using flux tower data over 82 sites in North America. The LSTM model with differentiated plant function types (PFTs) demonstrates the capability to explain 79.19% (R2= 0.79) of the monthly variations in NEE within the testing set, with RMSE and Mean Absolute Error values of 0.89 and 0.57 g C m−2d−1respectively (r= 0.89, p< 0.001). Moreover, the LSTM model performed robustly in predicting cross‐site variability, with 67.19% of the sites that can be predicted by both LSTM models with and without distinguished PFTs showing improved predictive ability. Most importantly, the IAV of predicted NEE highly correlated with that in flux observations (r= 0.81, p< 0.001), clearly outperforming that by the random forest model (r= −0.21, p= 0.011). Among all nine PFTs, solar‐induced chlorophyll fluorescence, downward shortwave radiation, and leaf area index are the most important variables for explaining NEE variations, collectively accounting for approximately 54.01% in total. This study highlights the great potential of LSTM for improving carbon flux upscaling with multi‐source remote sensing data. Net ecosystem exchange (NEE) of CO2is a crucial process that regulates carbon exchange between terrestrial ecosystems and the atmosphere. Currently, the growing availability of flux tower measurements and the rapid development of artificial intelligence technology, have made the machine learning‐based data‐driven flux upscaling approach a popular way for estimating carbon fluxes over large scales. Several upscaling NEE data sets have been derived with different machine learning methods; however, the lack of representing memory effects of climate and environmental factors in the modeling remains an important source of uncertainty for NEE estimates. To address this issue, we constructed site‐level Long Short‐Term Memory (LSTM) training models by plant function types in North America to improve the simulation of monthly scale NEE and its interannual variations. The established LSTM model improves the prediction of the temporal variability of NEE, showing great potential for improving carbon flux upscaling. The Long Short‐Term Memory (LSTM) model with differentiated plant function types (PFTs) demonstrates the capability to explain 79.19% of the monthly variations in net ecosystem exchange (NEE)The LSTM model exhibited clear advantages over the RF model in capturing the interannual variations of NEEThe relative importance of feature variables for predicting monthly NEE dynamics across different PFTs in North America was quantified The Long Short‐Term Memory (LSTM) model with differentiated plant function types (PFTs) demonstrates the capability to explain 79.19% of the monthly variations in net ecosystem exchange (NEE) The LSTM model exhibited clear advantages over the RF model in capturing the interannual variations of NEE The relative importance of feature variables for predicting monthly NEE dynamics across different PFTs in North America was quantified