【作者】 张静；

【导师】 江朝晖；

【作者基本信息】 安徽农业大学 ， 计算机应用技术， 2015， 硕士

【摘要】 作物生长模型是精准农业、数字农业和智能农业的重要支撑,近年来模型研究呈现出内容细致化、目标多样化和应用具体化等发展趋势。基于小样本数据的作物生长模型具有低耗、灵活、高效、兼容性强等优点,能够弥补大型综合性模型在作物生理学研究、环境胁迫研究以及温室调控应用等方面存在的不足,但如何确保模型性能是小样本作物生长建模的难点和研究意义所在。本文以多种作物的光合生理-环境实测数据为对象,研究小样本作物生长建模的若干关键问题,探讨确保模型性能的可行方法。主要内容和结果包括:(1)监测数据的预处理研究。针对PTM-48A型作物光合生理-环境参数监测仪对草莓、黑豆、番茄、南瓜、黄瓜等作物的连续监测数据,采用聚类分析对数据集的时间维度进行分类,获得连续、全面、均匀的数据段,再针对聚类获得的符合要求的数据段进行异常检测,剔除干扰数据点。通过聚类分析及异常检测处理,为后续分析和建模提供了可靠的样本数据。(2)因子分析与选择研究。以作物的CO2交换速率及其环境影响因子的相互关系为对象,定量地研究、比较了数据建模中常用的两种多因子分析方法——相关分析及通径分析的性能。实验结果表明,与相关分析相比,通径分析能够解释因子间的直接和间接作用效果,分析过程更全面,而且能够消除变量间的多重共线性,变量选择更有效,更适宜小样本数据作物建模。(3)模型构建方法研究。研究、比较了两种代表性建模方法——回归模型和人工神经网络模型的性能。针对四种常见回归模型的实验表明,纯二次模型精度较高而复杂度较低,更适宜用作小样本数据作物建模;运用黄瓜的CO2交换率-环境参数,建立了性能良好的小样本GA-BP预测模型;与纯二次回归模型相比,GA-BP神经网络模型更适宜用于小样本数据的作物模型构建。更多还原

【Abstract】 Crop model is one of the important means of precision agriculture,digital agriculture and intelligent agriculture.In recent years,the research of the crop growth model shows a trend of meticulous content and diversified objects.The crop growth model for small sample size has many advantages in low consumption,flexibility,efficiency,compatibility and so on.Besides,it can also make up for deficiencies in crop physiology,environmental stress and greenhouse control,in comparison with the major integrated model.But currently,to ensure the performance of model is the chief difficulties and significant of the crop growth models for small sample size.The paper mainly studies some key problems of crop model for small sample size,using the photosynthetic physiology and environmental monitored data of many crops.The chief content and result are listed as bellow.(1)Research on the preprocessing of monitored data.This paper used crop photosynthetic physiology and environmental monitoring system named PTM-48 A for continuous non-destructive monitoring of crops included strawberry,black beans,tomato,pumpkin,and cucumber.It adopted the cluster analysis to classify the measured time data,in order to obtain a continuous,comprehensive and uniform time period.Then the anomaly detection of data segment which meets the requirements for clustering,to detect and remove the outlier.The preprocessing based on cluster analysis and anomaly detection could offer a reliable sample data for the further analysis and modeling.(2)Research on factor analysis and selection.The study compared the performances of two multiple statistical analysis methods usually used in modeling,correlation analysis and path analysis,adopting the relation between CO2 exchange rate and the environmental influence factors.The analysis results showed that,compared with the correlation analysis,the path analysis was more suitable for crop models of small sample size.On the one hand,it could analyze more comprehensively,as it explained the direct and indirect effects between factors.On the other hand,path analysis was more efficient in variable selection,as it could eliminate the multicollinearity effect among variables.(3)Research on methods of modeling.The paper analyzed and compared two representative methods,regression model and artificial neural network model,which are commonly used in model fitting and construction.The experimental results of four common regression models showed that the pure quadratic regression model was more suitable for crop modeling of small sample size because of its higher accuracy and less complexity.On this basis,the forecasting model of CO2 exchange rate of the cucumbers with the use of GA-BP neural network was established.Compared with the pure quadratic regression and BP neural network,the GA-BP neural network was the more suitable model for crop model of small samples.更多还原