【作者】 马静；

【导师】 郑永果；

【作者基本信息】 山东科技大学 ， 计算机应用技术， 2003， 硕士

【摘要】 90年代卫星遥感在全球和区域尺度土地覆盖研究与应用方面取得了突破性进展，土地利用／覆盖遥感研究的新方法不断出现。遥感图像的分类是遥感数据在土地资源分析及应用的第一步，如何解决多类别图像的识别并满足一定的精度，是遥感图像研究中的一个关键问题，具有十分重要的意义。 与传统统计方法的分类器相比较，人工神经网络法不需要预先假设样本空间的参数化统计分布，正在被越来越普遍的应用于遥感图像分类的研究。在神经网络的不同算法中，应用和研究最多的是反向传播人工神经网络模型(简称BP网络模型)。 反向传播神经网络方法可用于遥感图像分类。本文在对BP网络分析的基础上，为提高网络的收敛速度和性能，提出了自己的一些建议，通过设置并改变样本的训练强度加快了网络学习的速度。另外还构造了一个用于混合像素分类的神经网络，输出层节点为各典型地物类别所占的百分比。分类前对数据进行主成分分析，实现特征提取的目的。在实际分类时仅使用波谱信息有时是不够的，可以添加一些辅助信息，如高程、植被指数等进行基于知识的推理判断。使用这些方法对实验区的TM图像进行了分类实验研究，取得了较好的效果。更多还原

【Abstract】 At present, there is a far-reaching use of remote sensing (RS) imagine in the field of land use/land cover change (LUCC) research. The research on LUCC is a core for studies on the global changes. Also the research of spatial-temporal features of land use/land cover change is significantly important for better understanding land use/land cover change and environmental management for sustainable development. The techniques of classification are very important for land use/land cover change (LUCC),so LUCC study are on the basis of the image processing and classification system in the first. How to improve the accuracy of RS interpretation in order to promote the utility of RS technology is a urgent problem in RS application.Compared with conventional statistic classifier, the artificial neural network (ANN) has been developed and applied to remote sensing data classification problem, which doesn’t need suppose parameterized distribution of sample space in advance. ANN has complicated mapping capability. The back propagation neural network modal (BP model) is often been used.Back propagation neural network classifier can solve the problems existing in the traditional classifiers and has been gradually used in the classification of remote sensing image. In order to accelerate the training speed, an improved BP method is to be presented after analyzing the BP model. Through setting a training intensity, the training time is reduced. As to the mix pixels, we construct a BP neural network which the nodes of input layer are the bands of remote sensing and the nodes of output layer are percent of several kinds of object. Before classification, Karhunen-Loeve transform is carried through for features abstract. In addition the spectrum information, many ancillary geo-information such as NDVI and Dem is contributed to judge. Synthesizing these methods, we have the test. Compared with classification of MLC method, the results show the improved method has not only the highest accuracy but also the fastest speed of classification.更多还原