gwk-NN并行遥感分类器的设计与实现

发布时间：2018-06-26 04:39

本文选题：k-NN + 地统计模型　；参考：《南京大学》2013年硕士论文

【摘要】：随着卫星技术的发展,遥感影像的分辨率和数据量正在急剧增长,对遥感影像的处理技术提出了更高的要求。近几十年来越来越多的研究人员投入到遥感分类方法的研究,使得分类精度不断提高,应用领域日益广泛。特别是地统计方法的引入,使得不少研究人员开始注意到地物空间关系对遥感影像分类的影响。k-NN方法是常用的最近邻分类器,易于与地统计模型相结合,将地物空间关系特征引入该分类器可提高分类精度。与此同时,许多领域对海量遥感数据处理的时效性也提出了更高要求,高效并行化遥感影像处理是解决这一问题的有效途径。因此,开发高精度的并行遥感分类器是遥感研究领域的一个重要发展方向。本文根据地理学第一定律和并行计算思想,提出了gwk-NN并行遥感分类器。具体而言,在常用的传统k-NN分类器的基础上结合地统计模型,建立地理加权的gwk-NN分类器,并采用数据并行、任务并行和双重并行等三种模式,从算法和数据两方面对分类器进行了并行化。其中,数据并行采用对等式,任务并行采用主从式,双重并行则混合使用了对等式与主从式并行方式。在此基础上,以SPOT 5遥感影像作为实验数据,选择地物类别典型、空间连续性好的区域进行试验,对gwk-NN并行遥感分类器的分类精度和性能进行了测试。结果表明：(1)gwk-NN分类器优于k-NN、最大似然法(ML)、神经网络(NN)和支持向量机(SVM)等传统分类器,分类精度得到显著提高,噪声明显减少甚至消失,而且方法同样简单、易用；(2)gwk-NN并行分类器在双重并行模式下性能最好,实用性明显增强。在该模式单机环境下,分类器的加速比最大达到了6.59,并行效率为82.4%,明显优于数据并行和任务并行模式下的分类器性能。综上所述,本文在以下两方面有所创新：(1)将不同的地统计模型与k-NN分类器结合得到新型gwk-NN分类器,明显改善了传统k-NN分类器的分类精度；(2)将并行计算引入遥感影像分类,对gwk-NN分类器进行并行化改进,提高了gwk-NN分类器的分类性能。
[Abstract]:With the development of satellite technology, the resolution and data volume of remote sensing image are increasing rapidly. In recent decades, more and more researchers have devoted themselves to the research of remote sensing classification methods, which makes the classification accuracy improve and the application fields become more and more extensive. Especially with the introduction of geostatistical methods, many researchers have begun to notice the influence of spatial relationship of ground objects on classification of remote sensing images. The k-NN method is a commonly used nearest neighbor classifier and is easy to be combined with geostatistical models. The classification accuracy can be improved by introducing the spatial relation features of ground objects into the classifier. At the same time, many fields also put forward higher requirements for the timeliness of massive remote sensing data processing. Efficient parallel remote sensing image processing is an effective way to solve this problem. Therefore, the development of high precision parallel remote sensing classifier is an important development direction in remote sensing research field. Based on the first law of geography and the idea of parallel computing, a parallel remote sensing classifier of gwk-NN is presented in this paper. In particular, based on the traditional k-NN classifier, a geo-weighted gwk-NN classifier is established based on the geostatistical model, which adopts three models: data parallelism, task parallelism and double parallelism. The classifier is parallelized from algorithm and data. Among them, data parallelism is peer-to-peer, task parallelism is master-slave, and dual parallelism is a mixture of peer-to-peer and master-slave parallelism. On this basis, the spot 5 remote sensing image is used as experimental data to test the classification accuracy and performance of the gwk-NN parallel remote sensing classifier. The results show that: (1) gwk-NN classifier is superior to k-NN, maximum likelihood method (ML), neural network (NN), support vector machine (SVM) and other traditional classifiers, the classification accuracy is significantly improved, the noise is obviously reduced or even disappeared, and the method is as simple and easy to use; (2) the gwk-NN parallel classifier has the best performance in the dual parallel mode, and its practicability is obviously enhanced. The speedup ratio of the classifier is 6.59 and the parallel efficiency is 82.4, which is superior to the performance of the classifier in data parallel mode and task parallel mode. To sum up, this paper has some innovations in the following two aspects: (1) combining different geostatistical models with k-NN classifier, a new type of gwk-NN classifier is obtained, which obviously improves the classification accuracy of the traditional k-NN classifier; (2) parallel computing is introduced into remote sensing image classification. The parallel improvement of gwk-NN classifier improves the classification performance of gwk-NN classifier.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP751;P237

【参考文献】