基于BP神经网络的遥感影像分类研究

发布时间：2017-12-26 22:33

本文关键词：基于BP神经网络的遥感影像分类研究　出处：《东华理工大学》2014年硕士论文　论文类型：学位论文

【摘要】：遥感技术(Remote Sensing,简称RS)以其数据高时效、多综合的特点,已经成为地球观测一个重要技术手段。根据遥感技术获取的遥感影像具有多光谱、高分辨率、多时相的特点,通过对原始数据做进一步的处理,就能够发现事物变化规律,从而有效预测其未来趋势,因此遥感影像的研究越来越受到学者的关注,而遥感影像分类则一直是该领域的研究热点。人工神经网络是模拟人脑功能的一种网络结构,具有自适应、自组织、自学习的能力,能够实现信息的分布式存储、并行处理,因其本身是一个非线性系统,适合于解决复杂的非线性问题,诸如遥感影像分类。BP神经网络是一种误差反向传播的神经网络,它能够将学习误差反馈到隐含层,改变初始网络结构的权值和阈值,从而达到预期的学习目标。实践证明,BP神经网络能够极大地提高影像分类精度。然而,BP神经网络在应用过程中也存在一些问题:易限入局部极小值,网络收敛速度慢,隐含层神经元个数无法确定,无法妥善解决影像中存在的“同物异谱,同谱异物”问题。针对上述问题,本文以覆盖江苏江阴、靖江及其周边的LANDSAT-7影像为实验数据,将遗传算法与BP算法相结合,对BP神经网络的初始权值进行优化,同时获得隐含层最优神经元数目。为了解决影像中存在的“同物异谱,同谱异物”问题,将NDVI指数值作为影像特征,与影像的纹理特征、光谱信息相结合共同用于影像分类。最后采用改进后的BP神经网络对结合NDVI指数、纹理特征、光谱信息的数据进行分类,并将其结果与传统方法的结果进行比较。实验结果表明:改进后的BP神经网络的分类效果和精度有了明显的提高。
[Abstract]:Remote Sensing (RS) has become an important technical means for earth observation because of its high aging and multi comprehensive data. According to the remote sensing image remote sensing images with high resolution, multi spectral and multi temporal characteristics, through further processing of the original data, we can found the changes of things, so as to predict its future trend, so the research of remote sensing image is more and more attention of scholars, and the remote sensing image classification is always a hotspot the field. Artificial neural network is a kind of network structure to simulate the human brain function, adaptive, self-organizing and self-learning ability, can realize distributed information storage, parallel processing, because of its itself is a nonlinear system, suitable for solving complex nonlinear problems, such as remote sensing image classification. BP neural network is an error back propagation neural network. It can feedback learning errors to the hidden layer and change the weights and thresholds of the initial network structure, so as to achieve the desired learning objectives. It has been proved that the BP neural network can greatly improve the accuracy of image classification. However, BP neural network has some problems in the application process: easy to limit into the local minimum, slow convergence rate, the number of hidden neurons can not be determined, unable to properly solve the image in the different spectrums of the same spectral problem. To solve the above problems, we take the LANDSAT-7 images covering Jiangsu Jiangyin, Jingjiang and its surrounding area as experimental data, combine genetic algorithm with BP algorithm, optimize the initial weights of BP neural network, and get the optimal number of neurons in hidden layer at the same time. In order to solve the image in the different spectrums of same spectrum, NDVI refers to the value as image features, texture features, image and spectral information together for image classification. Finally, the improved BP neural network is used to classify the data combined with NDVI index, texture feature and spectral information, and the results are compared with the results of traditional methods. The experimental results show that the classification effect and accuracy of the improved BP neural network have been greatly improved.
【学位授予单位】：东华理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：P237

【共引文献】