【作者】 王小霞；

【导师】 李书琴；

【作者基本信息】 西北农林科技大学 ， 计算机应用技术， 2009， 硕士

【摘要】 烤烟作为一种重要的经济作物,施肥对烟叶的产量、产值和品质都起着决定性的作用。传统施肥模型往往建立在专家经验和大量田间试验的基础上,存在主观性强、试验周期长和通用性差等缺陷。为更好地指导烤烟精准施肥,解决肥料配制中的技术难题,本文在综合评价烤烟土壤肥力的基础上,分别建立了基于BP网络和最小二乘支持向量机的智能施肥推荐模型,设计开发了BP网络和最小二乘支持向量机两种施肥推荐平台,并结合GIS技术用于指导烤烟施肥。本文的主要研究内容如下:(1)烤烟土壤肥力适宜性评价研究。比较现有的土壤肥力评价方法,在专家指导下建立评价因素集,选用灰色关联分析和AHP-模糊综合评价两种方法,分别对研究区的烤烟土壤肥力进行评价。依据各地烤烟的历史产量水平,对两种方法的评价结果进行验证。结果表明,AHP-模糊综合评价法相比灰色关联分析法,评价结果更准确,评价过程更符合实际。(2)建立基于BP网络平台的烤烟施肥推荐模型。针对传统施肥模型的不足,采用附加动量法和自适应学习速率法改进BP算法,构建基于BP网络平台的施肥推荐模型。综合考虑土壤养分差异对施肥量的影响,以土壤养分含量和施肥方案作为输入,以评价目标,如烤烟产量、产值或品质作为输出,BP网络各层神经元数由用户动态设定。实验结果表明,该模型对产量预测的平均准确率为88.55%,施肥推荐结果与实际一致。(3)建立基于最小二乘支持向量机的烤烟施肥推荐模型。分析最小二乘支持向量机模型的优点和适用性,研究其基本原理,构建基于最小二乘支持向量机的智能施肥推荐模型。总结常见的核函数形式,选择径向基函数作为施肥推荐模型的核函数;对于模型中的两个待定参数,采用粒子群优化算法来确定。实验结果表明,基于最小二乘支持向量机的施肥推荐模型对产量的预测准确率高于BP网络模型,且特别适用于样本数量有限的情况。(4)基于GIS的烤烟智能施肥推荐平台的开发。将构建的智能模型用于烤烟施肥推荐,设计开发了BP网络和最小二乘支持向量机两种施肥推荐平台。在GIS支持下,建立空间数据库,关联土壤肥力评价结果,绘制了烤烟土壤肥力适宜性分布图;将其与施肥推荐方案集成,发布到Web页面。测试结果表明,所建平台能有效地实现烤烟智能施肥推荐,具有较高的通用性、直观性和实用价值。更多还原

【Abstract】 Tobacco is an important economic crop. Fertilization plays a decisive role on tobacco yield, output value and quality. Conventional fertilization models are often constructed on the basis of expert experience and plenty of experimentation. Most of them have many faults, such as strong subjectivity, long experimental period, weak generality, etc. In order to improve the ability of instructing tobacco fertilization accurately and solve the problems in fertilizer preparation. Two intelligent fertilization models are constructed after comprehensive evaluation of soil fertility for tobacco. One is based on Back- Propagation neural network, and the other is based on Least Squares Support Vector Machine. Accordingly, two intelligent fertilization recommended platform are developed.Finally, GIS is applied to instruct tobacco fertilization.The main content of this research include:(1) Study on comprehensive evaluation of soil fertility for tobacco plantation based on GIS. By analyzing and comparing staple methods on fertility evaluation of soil, the set of evaluation factors is built under expert direction, and two methods are selected to evaluate soil fertility of researchful areas, including Gray Correlation Analysis and AHP- Fuzzy Comprehensive Evaluation. According to the historical yield, the evaluation results of two methods are verified. Experimental results show that the accuracy of AHP- Fuzzy Comprehensive Evaluation is higher than Gray Correlation Analysis, and the former is more tally with the practice.(2) Construction of tobacco fertilization recommendation model based on BP neural network. In view of the faults of conventional fertilization models, normal BP algorithm is improved using methods of the affixation momentum and the self-adaptation learning rate. A fertilization recommendation model is constructed based on BP neural network platform. Take most factors into consideration, some soil nutrient content and fertilization plans are viewed as input vector, and evaluation target (yield or output value or quality) is viewed as output vector. The number of neuron nodes in every layer of BP neural network is determined by user dynamically. Experimental results show that the mean prediction accuracy of this model is 88.55%, and the recommendation results coincide with the practice. (3) Construction of tobacco fertilization recommendation model based on Least Squares Support Vector Machine(LS-SVM). By analyzing the merits and applicability of LS-SVM and studying on LS-SVM fundamental principle, an intelligent fertilization recommendation model is constructed based on LS-SVM. Comparison with familiar kernel function forms, Radial Basis Function is selected as the kernel function of this model. In this paper, Particle Swarm Optimization algorithm is used to optimize model parameters. Experimental results show that the mean predicting accuracy of the LS-SVM model is higher than BP neural network model. At the same time, the LS-SVM model is especially suitable for limited number of samples.(4) Development of intelligent fertilization recommendation platform for tobacco. We apply the two models constructed in this paper into recommend fertilization for tobacco and develop two fertilization platforms. We construct a spatial database, associating with inner properties and draw a fertility distribution map for tobacco plantation in GIS environment. The results of fertilization recommendation are integrated with the distribution map and released to web page. Testing results show that both platforms can recommend fertilization for tobacco effectively and has pretty good commonality and practical value.更多还原