Synergistic importance of memory and spatial neighbourhood effects in modelling net ecosystem productivity

• Joint memory–neighbourhood modelling improves NEP simulation accuracy . • Ignoring spatiotemporal dependencies leads to overall carbon sink overestimation . • Spatiotemporal modelling improves detection of responses to extreme events. Net ecosystem productivity (NEP) is a key indicator of terrestrial carbon balance, and elucidating its spatiotemporal patterns is essential for understanding global carbon cycle processes and response mechanisms under climate change. However, substantial uncertainty persists owing to lag, legacy, and spatial neighbourhood effects, necessitating the joint consideration of memory and neighbourhood information for accurate NEP simulation. We developed a convolutional long short-term memory (ConvLSTM) deep-learning model using global flux observations, multisource remote sensing, and environmental driver data to capture spatiotemporal dependencies in simulating global terrestrial NEP and quantitatively compared it with models that consider only memory (LSTM), only neighbourhoods (CNN), or neither (ANN) to assess the synergistic importance of memory and neighbourhood effects in NEP simulations. The results demonstrated that (1) ConvLSTM performed best (mean site-level validation R 2 = 0.72; RMSE = 0.86 g C m -2 d -1 ), with closer agreement to observations in long-term trends and interannual variation. LSTM (R 2 = 0.66; RMSE = 0.94 g C m -2 d -1 ) and CNN (R 2 = 0.67; RMSE = 0.93 g C m -2 d -1 ) results were intermediate, and the ANN was the worst (R 2 = 0.56; RMSE = 1.06 g C m -2 d -1 ). (2) The results from ConvLSTM indicate that global terrestrial NEP increased from 2001 to 2023 (0.052 Pg C yr -2 ), with marked regional differences: parts of the Amazon and Congo showed declining sink capacity, whereas eastern China and India strengthened. Ignoring both memory and neighbourhood effects overestimated sink strength and growth in highly productive regions, and memory effects dominated neighbourhood effects in shaping NEP spatial patterns. (3) Jointly modelling memory and neighbourhood information also improved the detection of drought legacy effects and NEP responses to ENSO events.