Long-term Simulation of Gross Primary Productivity and its Impact Evaluation Based on a Mechanism and Data Co-Driven Model

Terrestrial Gross Primary Productivity (GPP) serves as a fundamental metric in carbon cycle investigations, offering insights into the health and resilience of terrestrial ecosystems. Long-term, high-precision GPP data facilitates comprehensive investigations into ecosystem dynamics and sustainability under the impacts of climate change and human activities. Here, we utilized the Noah-MP model, which simultaneously assimilates soil moisture data and Solar-Induced Chlorophyll Fluorescence (SIF) data, to simulate GPP. However, the estimation of long-term (1982–2020) GPP in semi-arid regions was constrained by the lack of SIF data prior to 2000. Therefore, this study synergistically integrated a mechanistic model (Noah-MP) and a data-driven model (Unet model) to derive long-term GPP. The synergy between these two approaches addressed their respective limitations, achieving faster computational speeds and higher simulation accuracy. Results demonstrated that data assimilation techniques effectively enhanced the simulation accuracy of the mechanistic model, and the Unet model incorporating attention mechanisms (Attention-Unet model) better reconstructed the spatiotemporal GPP data from 1982 to 2020. Based on the simulations, ecological restoration projects significantly improved regional GPP, exhibiting distinct seasonality and zonality. However, influenced by climate change, the carbon sequestration capacity of natural vegetation zones experienced a marked decline in 2008, while human activities exerted significant positive impacts on agricultural areas (e.g., Hetao Irrigation District), offsetting some climate change effects in these regions. Although ecological restoration projects enhanced vegetation carbon sequestration at annual scales, this upward trend did not steadily persist with the advancement of ecological engineering initiatives.