【作者】 刘纯平；

【导师】 夏德深；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2002， 博士

【摘要】 近年来，多源遥感数据融合技术已经成为遥感领域的研究热点。面对遥感数据多元化这个现实，多源信息融合方法的研究，特别是针对处理多源信息中的不确定性和不精确性问题的快速算法的研究，成为现今遥感数据分析和处理的重要内容。因此如何利用遥感多源信息进行有效的快速分类在计算机辅助分析和识别系统中更有实用价值。本文针对遥感图像计算机自动分类中的问题，主要进行了以下几方面的研究工作。 1) 改进基本的Kohonen神经网络(KNN)学习率和核函数，并针对每个信息源对最终识别贡献的大小，提出加权与不加权两种融合规则。算法在模拟多源图像融合分类中得到有效的应用。实验结果表明，采用加权的融合方法，可以取得更好的分类效果，对混合像元的识别也有一定的改善。 2) 针对基于KNN融合方法在遥感图像处理中的不足，提出使用模糊技术与KNN相结合的模糊Kohonen神经网络(FKNN)融合方法。实验表明，该融合分类方法能够快速收敛，并能得到较好的分类结果。 3) 由于构造一个恰当的隶属函数等同于提取有效特征，提出了MFKNN1和MFKNN2两种改进方法。同时对MFKNN2中神经网络的权值修改还引进一个平滑因子，以保证权矢量在输出空间分布的平滑性。有效地改善了融合分类的效果和算法性能。 4) 基于模糊逻辑、神经网络和遗传计算相结合，提出了进化规划FKNN(EPFKNN)融合算法。实验结果表明，EPFKNN融合分类算法的性能和分类结果比较好，尤其对混合像元的分类，可以得到更好的结果。 5) 研究了基本DS证据理论(BDSET)和模糊DS证据理论(FDSET)两种决策层融合方法。根据遥感图像分类原理以及证据理论的特点，分别确定了BDSET方法和FDSET方法的基本概率分配函数及融合分类规则。实验结果表明基于BDSET和FDSET融合的分类方法比传统的非监督分类方法具有更好的分类效果，有效地提高了分类的精度。 6) 结合多源遥感信息融合分类的特点，在多级数据融合方法集成技术的基础上，初步形成一个对多源遥感信息融合分类的技术框架。将特征层融合的FKNN与决策层融合的FDSET相结合，提出一种新的融合方法。在多源遥感信息融合分类的应用中，可以利用FKNN的自学习和自适应功能，根据训练样本的学习而自动地获得知识，并将这一知识应用到决策层融合的FDSET融合方法中。实验结果表明，摘要博士论文这种基于FKNN和FDSET相结合的多源遥感信息融合分类器经过训练后，可应用于遥感图像的计算机自动分类，其分类精度明显高于其他传统的计算机自动分类方法。 总之，本文的工作在应用计算机辅助多源遥感信息融合分类的领域内作了有益的理论探讨和方法研究，对计算机多源遥感信息自动分类的进一步完善和发展起了重要的推动作用。更多还原

【Abstract】 In recent years multi-sources data fusion techniques have already been an international research hotspot in Remote Sensing. Get useful information from Remote Sensing image and another information source of large areas is a time-consuming process. Computer-aided classification has provided an alternate, effective method. The main drawback of traditional remote sensing information computer classification methods is its low precision. Improving classification accuracy is a key issue in Remote Sensing image classification. Since multi-sources data fusion technique can efficiently improve the accuracy and the ability of fault tolerances, especially can deal with the problem of uncertainty and inaccuracy, it will play an important role in the multi-sources data analysis and process of Remote Sensing at present. According to the problem in the development of computer Remote Sensing image auto-classification, this paper does more detailed research in the area of feature and decision fusion classification.Based on unsupervised classification, improve the original Kononen Neural Network (KNN) by modifying learning rate and kernel function. And developed two fusion rules based on the magnitude of providing recognition information of each information sources. One is weighted; another is not weighted. Experimental results show that it can obtain better classification, as well as the recognition results of mix pixel are improved at some degree.Although fusion method based on KNN can deal with multi-sources Remote Sensing information, results show that a fast fusion classification method for a great number of image and another information sources has not yet devised. Though some detailed analysis of Remote Sensing information, this paper developed a method of fusion classification, which improves accuracy and convergence, by integrated fuzzy technique and KNN.Since structuring a right membership function, that is also effectively extracted feature, this paper proposed two modified method MFKNN1 and MFKNN2. In addition, in order to guarantee the smoothness of weight vector distribution in output space, MFKNN2 method is introduced a smooth factor in the modification of connective weight coefficient between neuron. Simulative experimental result suggested that these two methods are effective in improving classification accuracy and capability of algorithm.By integrated fuzzy logic, neural network and genetic algorithm, a fusion algorithm ofevolution programming FKNN is presented. Based on learning sample, this method can automatically determined the node of output space and can find a global minimum. So EPFKNN method results in accuracy and quickness. For classifying mix-pixel, it shows that the capability is better.Another emphasis is the fusion method research of decision level in this thesis. In fact, two methods are researched. One is the Basic Dempster-Shafer evidence theory (BDSET); another is the fuzzy Dempster-Shafer evidence theory (FDSET). In evidence theory, it is an important that how to obtain Basic Probability Assignment (BPA) function. According to the classification theory of Remote Sensing image, BPA function and the rule of fusion classification are determined respectively. The experimental results show that these two classification methods of multi-sources information fusion can result in better accuracy than that of conventional unsupervised classification method.Discuss the method and technique about integrated multiple levels data fusion based on the characteristics of multi-sources Remote Sensing information fusion classification. A classification frame of multi-sources Remote Sensing information fusion is designed by using FKNN of feature level fusion and FDSET of decision level fusion. Based on the multi-sources Remote Sensing information available, the frame of classification can be used computer-aided automatic classification by training. The classification accuracy is distinctively superior to conventional classification method.In general, some theory discussion and method r更多还原