基于海量解译标志的遥感影像卷积神经网络分类研究

发布时间：2018-06-23 17:43

本文选题：解译标志 + 卷积神经网络　；参考：《中国地质大学(北京)》2017年硕士论文

【摘要】：如何快速从海量遥感数据中获取遥感信息,实现遥感影像智能解译,一直是遥感技术领域研究的重要课题。人工神经网络技术在遥感影像自动分类领域取得显著成果。通过对国内外研究现状分析可发现,许多研究均是以改善算法为手段,以提高分类精度为目的,且分类精度已达到90%以上,但这些研究均是针对某一实验区域、实验数据进行的研究,当研究区域和实验数据改变之后,可能达不到很好的分类精度,而且自动化程度也不够。本文从改变样本来源、控制样本数据质量和保证样本数据数量等方面进行研究,以期提高分类的自动化程度和精度。本文将深度学习理论与人工神经网络技术相结合,并将最为成功的卷积神经网络模型应用于高分辨率遥感影像的分类进行实验研究。深度学习的优势在于从大数据中自动进行特征的提取与学习,卷积神经网络通过局部区域感知和权重共享两大法宝大大减少参数个数,降低了样本训练难度。所采用的实验数据不是特定的某一区域,而是取自多幅遥感影像的解译标志数据。本研究的主要成果如下:(1)对人工神经网络和卷积神经网络在遥感影像自动分类中的应用现状做了整理和分析,介绍了人工神经网络遥感影像分类的理论基础,重点介绍了深度学习的优势及卷积神经网络的特点和模型结构;(2)基于Visual Studio 2015软件平台,采用C#程序设计语言和WPF界面框架完成卷积神经网络框架的实现;(3)以解译标志作为卷积神经网络的样本数据,解译标志具有数据量大和保证真值两个特点,通过控制解译标志的来源和规范化存储来保证解译标志的质量,通过批量生成解译标志功能保证解译标志的数量;(4)以解译标志生成的样本数据为实验数据,运用实现的卷积神经网络框架完成高分辨率遥感影像的分类实验,得出影像本身分辨率越高,地物样本的质量越好,训练样本的数量规模越大,测试准确率越高的结论,但是随着训练样本数量规模的不断增大,准确率增长速度变缓。最终本文的分类精度为86.17%,在可用范围之内,消除了只针对一幅影像进行研究的局限性,对提高遥感影像分类的自动化程度具有很重要的现实意义。
[Abstract]:How to quickly obtain remote sensing information from massive remote sensing data and realize intelligent interpretation of remote sensing images has always been an important subject in the field of remote sensing technology. The artificial neural network technology has made remarkable achievements in the field of remote sensing image automatic classification. Through the analysis of the present research situation at home and abroad, it can be found that many studies are aimed at improving the algorithm and improving the classification accuracy, and the classification accuracy has reached more than 90%, but these studies are aimed at a certain experimental area. When the experimental data and the experimental data change, it may not achieve a good classification accuracy, and the degree of automation is not enough. In order to improve the automation and accuracy of classification, this paper studies on changing the source of samples, controlling the quality of sample data and ensuring the quantity of sample data. In this paper, the theory of depth learning is combined with artificial neural network technology, and the most successful convolution neural network model is applied to the classification of high resolution remote sensing images. The advantage of deep learning lies in the automatic feature extraction and learning from big data. Convolution neural network greatly reduces the number of parameters and reduces the difficulty of sample training by means of local region perception and weight sharing. The experimental data is not a specific region, but is derived from the interpretation of multiple remote sensing images. The main achievements of this study are as follows: (1) the application status of artificial neural network and convolutional neural network in remote sensing image automatic classification is summarized and analyzed, and the theoretical basis of artificial neural network remote sensing image classification is introduced. The advantages of deep learning and the characteristics and model structure of convolution neural network are introduced emphatically. (2) based on Visual Studio 2015 software platform, Using C # programming language and WPF interface framework to complete the implementation of the convolution neural network framework. (3) using the deciphering mark as the sample data of the convolution neural network, the deciphering symbol has the characteristics of large amount of data and ensuring true value. The quality of deciphering marks is guaranteed by controlling the source and canonical storage of deciphering marks, and the number of deciphering marks is guaranteed by generating the function of interpreting signs in batches. (4) the sample data generated by interpreting marks are taken as experimental data. The classification experiment of high resolution remote sensing image is completed by using the realized convolution neural network framework. The conclusion is that the higher the resolution of the image itself, the better the quality of the ground object sample, the larger the number of training samples and the higher the test accuracy. But as the number of training samples increases, the rate of accuracy increases slowly. Finally, the classification accuracy of this paper is 86.17. Within the available range, the limitation of only one image is eliminated, and it is of great practical significance to improve the automation of remote sensing image classification.
【学位授予单位】：中国地质大学(北京)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P237

【相似文献】