【作者】 罗凯；

【导师】 冯仲科；

【作者基本信息】 北京林业大学 ， 地图学与地理信息系统， 2008， 硕士

【副题名】以北京昌平区为例

【摘要】 随着遥感技术的不断成熟,土地遥感分类技术已成为获取土地变化信息的主要技术手段之一。本论文在系统介绍了国内外土地遥感分类研究中的主要方法和成果的基础上,重点介绍了当前发展较快的人工神经网络方法的特点以及它在遥感分类中的应用。研究以北京市昌平区为例,应用神经网络中的自组织特征映射网络和专家分类系统相结合对该区的土地利用类型进行分类。首先,对研究区影像数据的地类模数进行研究分析。由于遥感影像中地类往往会存在“异物同谱”、“同物异谱”的现象,所以研究的土地利用类型数目一般不等同于地类的模数。研究采用的自组织特征映射网络属于无导师学习的训练方法,因此网络的输出结果要尽可能地接近地类的模数。然后采用专家分类系统并导入研究区的数字高程数据对分类结果进行再次分类,最后得到研究区的土地类型分类结果。其次,对分类的结果进行精度评价和对比分析。采用类型精度评价法对研究的分类结果进行评价,并与传统的无监分类结果进行比较分析,结果表明该神经网络分类方法比传统分类方法在分类精度上有了较大的提高。最后针对研究中存在的问题,提出如何进一步提高分类精度的方法和建议。本论文针对现在分类方法往往存在单一性的特点,把自组织特征映射网络和专家分类系统结合起来,实现无监分类和先验知识相结合,有效地提高了分类的精度。研究表明根据不同分类方法的各自特点进行互补性结合并加入其它地理辅助数据,能有效地提高遥感分类方法的精度。更多还原

【Abstract】 On the basis of introducing the main methods and achievements in the field of land types classification based on remote sensing image in China and in abroad, the paper mainly introduces the features of artificial neural network and application in classification of remote sening.The paper combined the self-organizing feature maps with expert classification to classify the land use type. Because of same object with different spectra and different objects with same spectrum, firstly the study analysis the modulus of object. The self-organizing feature maps(SOFM) is an Unsupervised learning classification method, so the number of nodes of output layer should equaled approximately the modulus of object when we applicate SOFM. Then the study applicate expert classifier to classify again.Secondly, Taking into account the effect of sample randomnes and the feature of ground object, equal sample points were drawn in different land use types of the result to evaluate the classification precision. The result of evaluation shows that SOFM could improve the accuracy the classification greatly than the traditional unsupervised classification.Lastly, the study give some methods and suggestions to improve the accuracy of classification. Due to different classification method have the feature of uniformity, the study conbime self-organizing feature maps with expert classification to realize the joint of Unsupervised learning and previous experience.更多还原