【作者】 秦钟；

【导师】 胡秉民； 于强；

【作者基本信息】 浙江大学 ， 生态学， 2005， 博士

【摘要】 陆地生态系统碳、水循环的研究是当前全球变化研究领域的核心问题。将大量的、多尺度的生态系统-大气问物质与能最交换的试验和观测与描述这一过程的各种模型相结合是碳、水循环研究所普遍采用的方法。以涡度相关技术为主要的观测手段的全球通量观测网络（FluxNet）可对生态系统与大气间的CO₂、水分和能最交换过程进行长期的、连续的微气象观测,为进一步研究生态系统物质与能量传输的机理和过程以及碳源/汇时空格局等的提供重要的数据信息.同时为各种碳、水循环模型的参数确定、模拟效果检验提供检验标准。 20世纪90年代以来,国际上在对生态系统碳、水循环的关键过程及其对各种环境因素的响应、生态系统能量与通量的模型模拟等方面己取得了较大的进展,但研究对象多集中于自然生态系统（如森林、草地、湿地等）,农田作为一种受人类活动直接影响的生态系统在FluxNet的研究中并不具有代表性。我国的通量观测与研究启动较晚,通量观测积累的数据有限,需要在继续进行连续通量观测的同时,分析典型生态系统CO₂和水热通量以及相关生态环境要素时空变化特征及其动力学机制,并在不断修正和完善现有模型的基础上,探索新的方法或手段,寻找和构建使用便捷灵活、模拟效果良好且实用性强的通量模拟模型。 中科院禹城农业综合实验站是目前中国陆地生态系统通量观测研究网络（ChinaFlux）唯一的针对农田系统碳、水循环研究的通量观测站点。本研究以该站的涡度相关观测数据为基础,对2002年11月至2003年10月期间农田（冬小麦与夏玉米轮作）与近地层大气间的能量和物质交换特征进行了详细分析,并采用多种模拟技术对CO₂和水汽通量进行了模拟和对比,在以下几方面取得了新进展:1)应用涡度相关的实测数据,对基于过程的综合性模型RZ-SHAW模型模拟作物逐时水、热通量和逐日蒸散量的功能进行了验证;2)采用小波变换方法对农田生态系统的CO₂和水汽交换通量的多尺度特征进行了分析:3)提出了应用神经网络技术进行通量模拟的思路,并在对农田生态系统的CO₂和水汽交换通量模拟巾取得了较好的效果:4)将支持向量机技术引入农田生态系统的CO₂和水汽交换通量模拟中,并证明其模拟效果优于神经网络:5)将Bayesian证据框架的自相关确定方法应用于神经网络建模过程中,将小同的环境因素按其对CO₂和水汽通量影响进行了排序. 本研究在对该站农田生态系统水热、CO₂通量特征分析及模拟的过程中,得出了以下几点结论:更多还原

【Abstract】 During recent decades, carbon and water cycles of terrestrial ecosystems has been the focus in the study of global change. Central to this research is the combination of observations of ecosystem-atmosphere mass and energy exchange with the associated models, which has been adopted commonly as a useful method by the sites worldwide. Long-term, continuous micrometeorological measurements with the eddy covariance technique provide the information for the investigation of ecosystem mass and energy transfer processes and the regulating mechanisms as well as the spatial and temporal pattern of carbon source/sink. Moreover, it allows for direct fluxes measurements at the time and spatial scales required to evaluating or parameterizing land-surface properties of a process-based ecosystem exchange models and the model validation.A better understanding of critical processes in water and carbon cycles and how CO₂, water and energy fluxes may respond to biotic and abiotic factors and a progress on their simulation have been achieved since 1990s when the Worldwide CO₂ flux measurements network （FluxNet） was formed. The majority of the studies focus on natural ecosystems （e.g. forest, grassland and wetlands） while studies on agricultural and other human manipulated systems are currently under represented in the FluxNet community. China terrestrial ecosystems fluxes research （ChinaFlux） was formed until 2002, monitors of CO₂, water and energy fluxes need to be carried continuously for the exploring the biophysical controls on typical ecosystems and their roles in the global carbon and water balance, great strength have to be made at the same time not only in the revise and perfectness of the original fluxes model, bout also in the establishment of the models more applicable, flexible and convenient for the success simulation of fluxes.Yucheng Comprehensive Experiment Station in the North China Plain, the only agricultural ecosystem station of ChinaFlux operated CO₂, water and energy fluxes measurements with the eddy covariance technique above a winter wheat and summer maize rotation field. Data sets obtained from November 1, 2002 [day of year （DOY） 305]] to October 20,2003 （DOY293） were used to characterize the variability of CO₂ flux and surface energy balance components and to examine how they response to modulations of environmental variables in this study, then several strategies were used to simulate carbon and water fluxes and their prosperities and modeling abilities for different circumstances werecompared too. In doing this, some new ideas emerged as follows:1) Validated the ability of a process-based model termed RZ-SHAW to estimate ET and energy fluxes for summer maize field with the eddy covariance measurements; 2) Oscillation features of the cropland biosphere-atmosphere carbon and water vapor fluxes profiles over a wide range of time scales were probed with the wavelet transform technique; 3) Three-layer back-propagation neural networks were developed and applied to explore their capability in modeling water vapor and carbon dioxide fluxes exchange between the surface of a summer maize field and atmosphere; 4) The least squares support vector machines （LS-SVMs） method was introduced to simulate the water and carbon fluxes. The result was compared with the neural networks operated on the same data set in addition; 5) An automatic relevance determination （ARD） integrated with the Bayesian framework was used to assess the relative importance of the input environment variables, which were believed to have constraints on the dynamics of water and carbon exchange. The major issues presented in this study include:1) During the whole study period, net radiation captured by the cropland was consumed in the form of soil heat flux （G）, latent （LE） and sensible heat fluxes （Hs）. Among them, the largest part of energy was routed to surface water vapor exchange and the smallest part went to the heating of soil. Crop leaf area index and soil moisture influenced the pattern of the energy dissipation was strongly in two aspects. Firstly, the switch of energy partitioning was associated with leaf emergence （from Hs to LE dominated） and senescence （from LE back to Hs dominated）; Secondly, the proportion of latent heat flux to the available energy （Ra） increased from 48.6% in the winter wheat growth period to 62.6% in the summer maize period owing to the frequent precipitation and more sufficient soil water content with the onset of the summer monsoon;2) The results of wavelet transform on the data sets of water vapor and carbon dioxide （Fc）fluxes obtained during the summer maize growth period revealed that both Fc and LE were characterized by high wavelet coefficients intermitted by low values alternatively during the time scale of 60-1 lOd; Peak level wavelet coefficients for Fc and LE occurred contemporary around September 26 [day after sowing （DAS 105）]; For 64d and HOd scales, carbon flux signal showed almost the same pattern of "U" with a hollow ranged from DAS37-DAS85 and two peaks on both ends, while water vapor flux signal during the same time scale with an inverted "U"; The evolution of leaf area index （LAI） of the summer maize had coherences with the patterns of wavelet coefficients for carbon and water fluxes in 1 lOd under the influences of environmental variables（Meteoro!ogical factors and soil water content）;3) Daily evapotranspiration （ET） of the summer maize predicted by RZ-SHAW had a good agreement with that measured by eddy covariance that the square of the linear regression coefficient is 0.83, total water vapor flux estimated was 396.2 mm, about 2.73% greater than the measured value. Besides. RZ-SHAW model had a good performance in modeling hourly energy budget components. Forinstance, indices of agreement （IA）for hourly LE simulated and measured were above 0.75, root mean square error （RMSE） were no more than 1.0;4) Three-layer back-propagation neural networks （BPNN） could be an interesting and viable alternative method for modeling surface-biosphere fluxes exchange without using detailed physiological information or specific parameters of the plant. However, being based on experienced learning theory, they have a few drawbacks such as non-convex training with multiple local minima, dependence on quantity and quality of training data set, the choice of the number of hidden units.An improved SVMs model termed least squares support vector machines （LS-SVMs）, which developed on the basis of statistical learning theory, could be used to model surface fluxes exchange without restrictive assumptions. Compared with the results of the artificial neural networks, it has stronger learning ability, better generalization ability and is less dependent on the size of training dataset.5) Mechanistic models such as RZ-SHAW in this study provided detailed information on the biophysical, eco-physiological and biochemical processes for the simulation of mass and energy exchange between ecosystem and atmosphere. Long-term experiments should be carried for calibration and specification of the parameters in the model, which makes fluxes estimation complex and carbon or water budget construction relatively inconvenient. Artificial neural networks （ANNs） and SVMs are interesting and viable alternative routines for modeling fluxes because they could extract underlying relationships between system inputs and outputs that are often very difficult to model with a mechanistic approach. However, the use of them should not diminish the role of mechanistic models, since these two methods are generally only superior in situations where the underlying system fundamentals are poorly understood or the parameters are difficult to obtain.6) The ARD results demonstrated that among several input variables for the neural network, vapor pressure deficient （VPD） and soil water content （W） had most influential effects on surface water vapor flux exchange while photosynthetically active radiation （PAR） and air temperature （T） were the key physical driving factors for carbon flux among several input variables. PAR, VPD, T and LAI were primary factors regulating both water vapor and carbon dioxide fluxes.Multi-years measurements of CO2, water and energy fluxes above the agri-ecosystem will be stuck and integrated carbon and water cycles simulation models will be developed for the purpose of a good knowledge of mass and energy transfer processes and their feedback mechanistic, and for an accurate estimation of the size of regional carbon source/pool in average from cropland and an efficient usage and regulation of water resources by upscaling of these site-based models. Impacts of land tillage, management measures and the crop species on both annual carbon fixation/release and water consumption will be evaluated in addition in future study.更多还原