陆地生态系统通过增加固碳和减少温室气体排放可以提升气候变化减缓能力，这种“基于自然的气候解决方案”(natural climate solutions，NCS)被认为是生态系统自然碳汇之外的额外潜力，对于实现国家“碳中和”目标至关重要。由于生态系统呼吸模型模拟精度存在较大差异，净生态系统碳通量(Net Ecosystem Exchange，NEE)模拟尚未考虑时间序列信息和空间异质性的影响，导致不同陆地生态系统碳通量模拟过程中存在较大的不确定性。本文基于传统经验呼吸模型、时间序列和空间信息的深度学习方法模拟了典型陆地生态系统碳通量。本文使用六种传统经验模型模拟了蒙古高原四种草地生态系统呼吸，对比不同模型在不同生态系统中的模拟结果，探究了影响草地生态系统呼吸的关键影响因素。使用两种基于时间序列的循环神经网络模型(门控递归单元Gated Recurrent Unit，GRU和长短期记忆模型Long-short Term Memory，LSTM)模拟净生态系统碳通量，在不同的季节尺度上分析了环境因子对碳通量变化的影响，揭示了生物物理因素对生态系统碳通量影响的滞后效应。基于通量足迹分析模型与遥感影像，结合图神经网络模型模拟生态系统碳通量，并分析不同的因子变换组合对碳通量模拟结果的影响，为弥补涡度相关碳通量监测的下垫面异质性对监测精度的影响提供重要理论支撑。主要研究内容和结论如下：
(1)六种传统经验模型模拟蒙古高原草地生态系统呼吸的效果分析
基于2014-2018年蒙古高原四种草地生态系统(灌丛、荒漠草原、典型草原和草甸草原)的通量观测数据，采用六种传统经验模型(Arrhenius、Logistic、Gamma、Martin、Concilio和Time series模型)来模拟生态系统呼吸(Ecosystem Respiration，Reco)。结果表明，六种传统经验模型的模拟精度为50%-80%。土壤温度和土壤湿度是影响草地生态系统呼吸变化的两个关键因素。以土壤温度和含水量驱动的模型在灌丛、荒漠草原、典型草原和草甸草原上Reco模拟精度较高，其中Martin模型表现最好(R2为0.57-0.81；RMSE为0.02-0.03)。在土壤湿度较高和水分胁迫较弱的草甸草原，六种Reco模型的模拟结果的R2为0.79-0.81。当土壤温度超过一定阈值时，比如灌丛约为19.62℃、荒漠草原约为16.05℃、典型草原约为16.92℃、草甸草原约为15.03℃，生态系统呼吸与土壤温度之间的指数关系不再存在，这通常与土壤湿度较低有关。除草甸草原外，Time series模型在灌丛、荒漠草原和典型草原生态系统中并未提高生态系统呼吸模拟的精度。生态系统呼吸模型的选择因生态系统类型及水文条件等而存在差异，不同类型草地生态系统呼吸模拟时模型的选择需要因地制宜。
(2)基于涡度相关和循环神经网络模型的不同季节尺度上碳通量模拟精度评估
基于2009-2020年美国凯洛格生物站(Kellogg Biological Station，KBS)七个站点(两个免耕玉米站点 CRP-C和AGR-C、两个柳枝稷站点CRP-Sw和AGR-Sw、两个恢复草地站点CRP-Pr和AGR-Pr以及一个参考站点CRP-Ref)的12年连续观测的通量数据模拟半小时CO2净生态系统交换(NEE)，训练了门控递归单元(GRU)模型，分析了不同季节尺度上(全年、生长季和非生长季)净生态系统碳通量的模拟情况，揭示了生物物理因素对NEE变化的相对重要性。结果表明，GRU模型在NEE模拟方面表现出较好的效果。在全年、生长季和非生长季三个季节尺度上模拟NEE的R2分别为0.89-0.93、0.85-0.9和0.61-0.85，比长短期记忆(LSTM)模型的模拟精度高1.17%-9.32%，运行时间短6%。入射辐射变量对NEE变化的贡献度最大，单个因素的最大贡献度为0.66-0.90。在非生长季，贡献度相对较高的变量包括入射短波辐射、年积日、温度、莫宁-奥布霍夫稳定度、风向和土壤含水量，其贡献度因生态系统类型和建模尺度而异。基于涡度相关和循环神经网络模型的净生态系统碳通量模拟为从时间序列角度模拟生态系统碳通量提供了新思路，为模型输入变量的选择提供了参考。
(3)基于空间信息和图神经网络模型的生态系统碳通量模拟与分析
基于2020-2022年中国南方红树林恢复区的通量观测数据，结合通量足迹分析模型与遥感影像数据，使用三种经典的图神经网络模型(GCN、GIN和DeeperGCN)模拟在不同空间分辨率下(10m、20m和30m)模拟生态系统碳通量，对比分析了20m空间分辨率下DeeperGCN模型与随机森林模型对NEE的模拟效果。结果表明，在20m空间分辨率下红树林恢复区NEE的模拟效果最好，并且DeeperGCN模型的模拟效果优于GCN和GIN模型；在20m空间分辨率下随机森林的模拟效果(R2=0.48)优于DeeperGCN模型(R2=0.45)。使用DeeperGCN模型对随机森林模型的模拟结果进行修正后，其模拟精度提高了3%。为了探究不同因子变换组合对模拟结果的影响，对环境因子进行了9种不同的变换组合，结果表明，分别将土壤温度和VPD变换为\ln{\left(TS\right)}和\sqrt{VPD}时，模型的模拟效果最好，精度提升了9%，且RMSE和MAE两种指标分别降低了0.17和0.19。通过结合植被生理过程数据与图神经网络模型，从空间信息角度模拟生态系统碳通量，并分析了不同因子变换组合对模拟结果的影响，为生态系统碳通量的模拟提供了新的方法以及变量选择的依据。
本文基于传统经验模型、循环神经网络模型以及图神经网络模型，分别从时间序列和空间信息的角度模拟了生态系统碳通量(Reco和NEE)，探究了环境因子对生态系统碳通量的影响，为模拟不同生态系统碳通量时模型以及影响因素的选择提供了依据，为从时间序列和空间信息的角度降低生态系统碳通量模拟的不确定性奠定了基础。

Terrestrial ecosystems can enhance climate change mitigation by increasing carbon sequestration and reducing greenhouse gas emissions, and this "nature-based climate solution" (NCS) is considered to be an additional potential to the natural carbon sinks of ecosystems, which is essential for achieving national "carbon neutrality" goals. Because the simulation accuracy of ecosystem respiration models is quite different, and the impact of time series information and spatial heterogeneity has not been taken into account in the simulation of net ecosystem exchange (NEE), there is a large uncertainty in the simulation of carbon flux in different terrestrial ecosystems. In this paper, carbon fluxes of typical terrestrial ecosystems were simulated based on traditional empirical respiration models, deep learning methods of time series and spatial information. Based on six traditional empirical models to simulate the respiration of four grassland ecosystems on the Mongolian plateau, we compared the simulation results of different models in different ecosystems and explored the key influencing factors affecting the respiration of grassland ecosystems. Two recurrent neural network models (Gated Recurrent Unit, GRU and Long-short Term Memory, LSTM) were used to simulate NEE, and the effects of environmental factors on NEE were analyzed at different time scales, revealing the lagged effects of biophysical factors on NEE. The lagged effect of biophysical factors on ecosystem carbon fluxes was revealed. Based on the flux footprint analysis model and remote sensing images, we combined the graph neural network model to simulate NEE, and analyze the effects of different combinations of factor transformations on simulation results, which provides important theoretical support to compensate for the influence of heterogeneity on the monitoring accuracy of NEE by eddy covariance. The conclusions are as follows:
 (1) Simulation of grassland ecosystem respiration in Mongolian Plateau based on six traditional empirical models
Performances of six empirical respiration models, including Arrhenius, logistic, Gamma, Martin, Concilio, and time series model were evaluated against measured ecosystem respiration during 2014-2018 in four grassland ecosystems on the Mongolian Plateau: shrubland, dry steppe, typical steppe, and meadow ecosystems. These models achieved good performance for about 50%-80% of the simulations. Both soil temperature and soil moisture played huge roles in simulating ecosystem respiration. Soil moisture-included model, especially Martin model, is more suitable for accurate prediction of ecosystem respiration (R2=0.57-0.81; RMSE=0.02-0.03). In meadow grasslands with high soil moisture and weak water stress, the simulation results of the six Reco models have R2 values of 0.79-0.81. When the soil temperature exceeds a certain threshold, such as ~19.62℃ at shrubland, ~16.05℃ at dry steppe, ~16.92℃ at temperate steppe, and ~15.03℃ at meadow, the exponential relationship between ecosystem respiration and soil temperature no longer exists, which is usually related to lower soil moisture. The Time series model did not improve the accuracy of ecosystem respiration simulation in shrubland, dry steppe, and typical steppe ecosystems. The selection of ecosystem respiration models varies depending on ecosystem types and hydrological conditions. The results of this study provide a certain basis for the selection of models for simulating respiration in different types of grassland ecosystems.
(2) Evaluation of NEE simulation accuracy at different time scales based on eddy covariance and recurrent neural network models
The gated recurrent unit (GRU) model was trained for simulating half-hourly net ecosystem exchange of CO2 (NEE) using 12-year continuous flux data (2009-2020) from seven experimental bioenergy crops in southwest Michigan. The fields were historically managed either as corn-soybean rotation agricultural (AGR) or Conservation Reserve Program (CRP) lands and planted to no-till corn (AGR-C and CRP-C), restored-prairie (AGR-Pr and CPR-Pr), switchgrass (AGR-Sw and CRP-Sw) and smooth brome grass (CRP-Ref). We compared NEE simulations for the entire year, growing season, and non-growing season datasets, and analyzed relative importance of biophysical variables on variations of NEE. GRU performed well in simulating NEE, with R2 of 0.89–0.93, 0.85–0.9, and 0.61–0.85 for the annual, growing season, and non-growing season datasets, respectively, which were 1.17%-9.32% higher and 6% of run-time shorter than those of long short-term memory (LSTM). Radiation parameters contributed most to the variability of NEE, with maximum unique contribution of 0.66-0.90. Contributions of individual variables appear more complex during the non-growing season but include incoming shortwave radiation, day of year, temperature, Monin-Obukhov stability, wind direction, and soil water content. The six most important forcing variables are consistent with our current understanding, although their ranked importance varied by ecosystem type and modeling scale. These results bring renewed insights in modeling observations at seasonal scales and provide guidance for variable selection to scale up measurements from one site to multiple land-cover types across the landscape.
 (3) Simulation and analysis of NEE based on spatial imformation and graph neural networks
Based on flux observation data of mangrove restoration areas from 2020 to 2022, and combined with flux footprint models and remote sensing data, three classic graph neural network models (GCN, GIN, and DeeperGCN) were used to simulate ecosystem carbon flux at different spatial resolutions (10m, 20m, and 30m). The results show that the simulation result of DeeperGCN model is better than GCN and GIN models at a spatial resolution of 20m. And the simulation effect of random forest (R2=0.48) is better than that of DeeperGCN model (R2=0.45). The simulation accuracy was improved by 3% after using the DeeperGCN model to correct the simulation results of the random forest model. To investigate the effects of different combinations of factor transformations on the simulation results, nine different combinations of transformations of environmental factors were performed, and the results showed that the accuracy of the model improved by 9% when transforming soil temperature and VPD into \ln{\left(TS\right)} and \sqrt{VPD}, and the two indicators of RMSE and MAE decreased by 0.17 and 0.19, respectively. By combining vegetation physiological process data and graphical neural network model, we simulated ecosystem carbon fluxes from a spatial perspective and analyzed the effects of different combinations of factor transformations on the simulation results, which provided a new method and a basis for variable selection for the simulation of ecosystem carbon fluxes.
By using traditional empirical models, recurrent neural network models, and graph neural network models, ecosystem carbon fluxes (Reco and NEE) were simulated from the perspectives of time series and spatial series, respectively. The impact of environmental factors on ecosystem carbon flux simulation was explored, providing a basis for selecting models and influencing factors for simulating different ecosystem carbon fluxes. This lays the foundation for reducing the uncertainty of ecosystem carbon flux simulation from a spatiotemporal perspective.