【作者】 徐浩；

【导师】 曹卫星；

【作者基本信息】 南京农业大学 ， 农业信息学， 2020， 博士

【摘要】 作物区域光温生产潜力模拟可为制定农业规划及决策提供重要依据。目前,基于过程的作物生长模型已广泛应用于区域光温生产潜力模拟,但是作物生长模型一般是在较为均匀的站点或田块尺度上建立和验证,而区域的自然环境要素存在空间异质性,作物生长模型区域应用存在尺度不一致的问题。因此,如何建立作物生长模型区域升尺度方法是急需解决的关键问题。本研究以中国冬麦区为研究对象,通过地理信息系统(GIS)与小麦生长模型WheatGrow耦合,探讨生长模型区域化应用的关键空间技术问题,建立WheatGrow模型空间应用升尺度连接方法,构建小麦区域光温生产潜力模拟预测软件,从而实现小麦生长模型的区域化应用,对完善数字农作平台及决策支持系统具有重要意义,同时可为识别作物产量区域限制因子,探究不同区域产量提升途径提供有力工具。通过构建6个嵌套空间分辨率的气象数据尺度序列,利用空间插值将气象站点数据插值成对应空间分辨率的栅格数据。然后利用WheatGrow模型模拟2000到2009年小麦区域光温生产潜力;设计尺度效应指数研究不同空间分辨率气象数据对区域光温生产潜力模拟的影响,并基于尺度效应指数阈值得出WheatGrow模型模拟区域光温生产潜力所需气象数据的适宜空间分辨率,同时研究利用地形的空间异质性获取WheatGrow模型光温生产潜力模拟所需气象数据适宜空间分辨率的可行性。结果表明:基于尺度效应指数能够得出WheatGrow模型区域光温生产潜力模拟所需气象数据适宜空间分辨率的空间分布,表现出地形空间异质性越强,气象数据空间异质性越强,区域光温生产潜力模拟所需气象数据适宜空间分辨率越高,反之亦然;虽然利用地形空间异质性可指导气象数据适宜空间分辨率选择,但在地形空间异质性较弱或地形空间异质性在小空间范围内很强的区域,单一半方差函数基台值无法准确反映模拟结果尺度效应,利用地形获取气象数据适宜空间分辨率存在一定局限性。基于中国冬麦区10年(2000-2009)538个气象站点小麦光温生产潜力模拟结果,采用空间随机抽样的方法设置17种不同站点数目,站点样本量数目从20到500个,间隔30,结合6种不同分区方案构建了 102种区域光温生产潜力模拟情景,并采用空间加权平均的方法计算区域尺度的小麦光温生产潜力。利用相对误差量化分区准确性,并通过计算基本空间单元内站点模拟产量的标准差量化分区稳定性,最终得出基于站点模拟结果获取区域光温生产潜力的适宜分区方案。结果表明,基于站点模拟结果升尺度获取区域光温生产潜力的方法受分区方案与站点分布的共同影响;当站点数目较少时,站点分布对不同分区方案区域光温生产潜力模拟影响显著,利用基于日照时数聚类的分区方案可以有效保证模拟精度;相反,当站点数目达到一定规模后,不同分区方案对区域光温生产潜力模拟结果影响较小。以格网化的WheatGrow模型模拟的小麦区域光温生产潜力为目标变量,利用区域上对光温生产潜力模拟有影响且容易获得的环境数据作为区域特征变量,基于多元线性回归、人工神经网络、随机森林与支持向量回归四种机器学习算法,构建基于特征变量的区域光温生产潜力模拟中间模型,达到在保证模拟精度的前提下,降低作物生长模型区域模拟应用中的数据需求,提高区域模拟效率的目的。结果表明:利用机器学习算法建立中间模型,易于获取的区域逐月气象数据可代替逐日气象数据,有效降低区域光温生产潜力模拟所需数据量,提高区域数据的可用性;四种机器学习算法建立的中间模型均可较好反映区域光温生产潜力均值,但随机森林建模性能最优,其次分别为人工神经网络、支持向量回归与多元线性回归;随机森林变量重要性与部分依赖图表明,经度、纬度、海拔与三月份最高温为区域光温生产潜力模拟的关键变量;机器学习模型预测能力受训练数据的影响,若预测数据范围超出训练数据的范围,则机器学习模型预测存在偏差。在分析WheatGrow模型输入数据、子模型依赖关系、计算流程与输出数据的基础上,利用Python语言对算法进行重构,建立格网化WheatGrow模型,实现了基于空间格网数据的区域光温生产潜力模拟。同时,采用基于消息传递接口(Message-Passing-Interface,MPI)的格网数据分区策略,采用并行计算方法,可以根据计算机的中央处理器(Central Processing Unit,CPU)核数和原始格网数据的大小,动态地将格网数据分割成一定数量的块。因此,区域光温生产潜力模拟的计算可以充分利用计算机的CPU容量,减少存储的物理内存的消耗。最后,采用GIS组件开发模式,在.NET平台下,结合C#和Python编程语言实现格网化WheatGrow模拟系统的开发,实现了区域光温生产潜力模拟及产量差计算等功能,为研究区域小麦光温生产潜力,评估气候变化对小麦生长影响,制定农业决策提供软件工具。更多还原

【Abstract】 Regional light temperature yield potential simulation can provide an important basis for agricultural planning and decision-making.At present,the process-based crop growth model has been widely used in regional light temperature yield potential simulation.However,the crop growth model is generally established and validated at the relatively uniform site or farm scales,while the spatial heterogeneity exists in the regional natural environment.So,the scale inconsistent exists in the regional application of the crop growth model.Therefore,establishing an up-scaling method for regional yield simulation is a key issue that needs to be solved urgently.In this study,we focused on China’s winter wheat region and studied the up-scaling methods in regional yield simulation.We investigated and determined the optimal spatial resolution and best zonation scheme in the regional application of the WheatGrowth model.We built a GIS-based platform for wheat regional productivity simulation and improved its efficiency.By constructing a nested sequence of spatial resolution,and we used the interpolation method to interpolate the site-specific meteorological data into the corresponding spatial resolution raster data.Then,the WheatGrow model was used to get the regional yield potential from 2000 to 2009.By constructing the scale effect index,we analyzed the influence of meteorological data with different spatial resolutions on the regional potential productivity simulation.We adopted the scale effect index threshold to determine the appropriate spatial resolution of the meteorological data required for the regional potential productivity simulation in the research area.Meanwhile,we analyzed the feasibility of using the spatial heterogeneity of the landform to obtain the appropriate spatial resolution of meteorological data required for the regional potential productivity simulation of WheatGrow.Results showed that we could obtain the spatial distribution of appropriate spatial resolution for the meteorological data required for the regional yield potential simulation of the WheatGrow model based on the scale effect index.Moreover,we could use landforms’ spatial heterogeneity to determine an appropriate spatial resolution for the meteorological data.However,in the regions where the landform’s spatial heterogeneity was relatively weak or relatively strong over a small range,using a single heterogeneity index derived from semivariograms cannot well reflect the scale effect of a simulation result.Limitations existed in obtaining an appropriate spatial resolution of meteorological data by landforms.Based on the simulated wheat yield potential at 538 sites throughout China’s main winter wheat production area from 2000 to 2009.Seventeen schemes were defined using spatial random sampling,and the number of sites per scheme ranged from 20 to 500 with an interval of 30.By combining six different zonation schemes,a total of 102 regional yield potential estimation scenarios were constructed.Then,the regional yield potential was calculated using the spatial weighted average method.The relative errors were calculated to quantitatively evaluate each zonation scheme’s impact on the accuracy of the simulation results.The standard deviations of the simulated yields for the sites in the basic spatial units were calculated and used to evaluate the effects of different zonation schemes on the simulation results’ stability.Finally,an appropriate zonation scheme for estimating the regional yield potential was obtained based on the site-specific simulation results.Results showed that the upscaled site-specific yield potential is affected by the zonation scheme and sites’ spatial distribution.The distribution of a small number of sites significantly affected the simulated regional yield potential under different zonation schemes,and the zonation scheme based on sunshine duration clustering zones could effectively guarantee the simulation accuracy at the regional scale.In contrast,the large number of sites had little effect on the regional yield potential simulation results under the different zonation schemes.We built regional yield potential simulation meta-models based on machine learning methods,including multivariate linear regression,multilayer perceptron,random forests,and support vector regression.The meta-models used regional yield potential simulated by the gridded WheatGrow model as target variables and the regional easily available environmental data,which affect regional yield potential as feature variables.Building the meta-models aims to minimize environment data requirements for regional yield potential simulation of crop growth model and improve the efficiency of regional simulation under the premise of ensuring simulation accuracy.Results showed that the proposed meta-models could use monthly weather data instead of daily weather data.These meta-models’applications could effectively minimize the data requirements in regional light temperature yield potential simulation,and improve regional data availability.The four machine learning models could well reflect the mean regional yield potential.However,random forest modeling performed best,followed by multilayer perceptron,support vector regression,and multiple linear regression.The variables’importance and partial dependence of random forest showed that longitude,latitude,altitude,and maximum temperature in March(Tmax-3)were the most critical variables in the regional light temperature yield potential simulation.Training data influence the prediction ability of machine learning model.If the prediction data range exceeded the range of training data,the machine learning model’s prediction would be biased.Based on the analysis of the WheatGrow model,including input data,sub-model dependency,calculation process,and output data,the model was reconstructed by Python language.A gridded WheatGrow model was established to realize the wheat regional productivity simulation based on spatial grid data.Meanwhile,with the grid data partition strategy based on Message Passing Interface(MPI),a parallel computing method was adopted.The method could segment the grid data into a certain number of blocks dynamically based on the number of the CPU cores and the original grid data size.So,the computation of the regional productivity simulation could take advantage of the full CPU capacity and reduce the consumption of the physical memory stored.Finally,we implemented the gridded WheatGrow simulation system based on the existed GIS components by using the Microsoft.Net developer platform with C#and Python programming language together,realized the functions of regional light temperature yield potential simulation and yield gap calculation.The assembled system could provide a software tool for evaluating climate change impacts on food security and making agricultural decisions.更多还原