【作者】 崔宝侠；

【导师】 徐心和；

【作者基本信息】 东北大学 ， 控制理论与控制工程， 2005， 博士

【摘要】 水环境评价是一个涉及多学科、多因素的复杂大系统，它的优劣将直接影响人们日常生活和经济的发展。决策支持系统(DSS)为人们提供了分析问题、构建模型、模拟决策过程和效果的决策环境，成为解决半结构化问题和非结构化问题的有效工具。 随着人工智能技术、地理信息系统的飞速发展，决策支持系统不仅能解决定量的问题，而且也能很好地处理不确定的、模糊的、定性的知识，从而辅助决策者做出科学的决策。针对当前辽宁省水资源短缺、污染严重的现状，为了提高水环境的管理水平、促进水资源的可持续发展，本文在设计基于GIS平台的水环境评价决策支持系统的过程中，主要完成了以下的研究工作。 1．针对遥感图像进行图像压缩的研究。首先分析了遥感图像的统计特性和相关性，并对各种码书设计算法进行分析研究。利用小波图像系数的空间分布特征以及它的多分辨率特点，应用遗传退火算法、有性繁殖遗传算法，提出了矢量量化方法在图像压缩中的应用。实验结果表明本文提出的码书设计方法可以在较大压缩比的前提下，得到更好的恢复图像质量，并有较快的收敛速度。 2．研究遥感图像分类的几种方法。结合遥感图像处理专业软件ERDAS IMAGINE，采用BP神经网络算法实现了非监督聚类算法和监督分类法的结合，使分类试验精度有了一定的提高。研究分析了自组织神经网络算法和学习矢量量化算法，提出了一种基于改进的SOFM算法和LVQ2算法的混合学习矢量量化算法。通过具体试验的实施，证明HLVQ算法不仅保持了数据的拓扑有序性，而且还进一步提高了分类精度。最后探讨了LVQ算法的主要缺点，在分析广义学习矢量量化算法的基础上，将GLVQ算法引入遥感影像分类中，并建立了分类模型。 3．模型库及模型库管理系统的开发与研制。为了适应对水环境评价和预测的需求，本文引进了大量的水质数学模型。在人工智能方面，除用于水质评价的模糊评价模型、用于水质预测的人工神经网络模型和基于遗传的神经网络模型外，本文将模糊神经网络应用于水环境影响预测方面，并取得了很好的预测效果。本文采用面向对象技术的方法，与数据库相结合，建立模型库管理子系统，有效地实现了对模型的增加、删除、修改，从而方便水环境管理者实现对模型的管理。更多还原

【Abstract】 Water environmental assessment is a complicated large-scale system which involves many subjects and factors, and its qualities will directly influence human’s daily life and economic development. Decision support system (DSS) provides a decision-making environment where we can analyze questions, construct models, and simulate the process and effects of decision; it has been an effective and powerful tool in solving some semi-structured and unstructured problems.With the rapid development of artificial intelligent technology and geography information system (GIS), decision support system not only solves quantitative problems but also deals well with the uncertain, fuzzy, qualitative information, and helps decision-makers to make sensible decisions. Facing the deficient water resource and seriousness of pollution, aiming at upgrading environmental management and promoting well-development of water resource, the researches about water environment assessment DSS design based on GIS are as follows:1. Study on remote sensing images compression technology. The statistical and relative characteristics of remote sensing images as well as some kinds of design methods in codebook are analyzed and studied. Depending on the spatial distribution characteristics and multi-resolution feature of wavelet coefficients, the vector quantization of codebook design in image compression has been push forward supported by genetic algorithm and annealing methods. The result shows that this codebook design method is able to achieve better image restoration quality, faster convergence speed, and better compression ratio than before.2. Researching on classification of remote sensing images. By means of remote sensing software, namely ERDAS IMAGING, the bond between unsupervised clustering algorithm and supervised classification has been strengthened through BP neural network, hence a higher precision of classification trials is got. Based on the self-organizing neural network and learning vector quantization algorithm, a hybrid learning vector quantization algorithm (HLVQ) combining the modified SOFM algorithm and the LVQ2 algorithm is proposed, and experimental results illustrate that HLVQ algorithm does well not only in retaining data topological structure but also in improving classification precision. In the end, the primary weaknesses of LVQ algorithm is discussed and GLVQ algorithm of remote sensing images classification is introduced, then classification models on the basis of the general learning vector更多还原