遥感影像分类的并行化及智能化研究

发布时间：2018-05-26 17:33

本文选题：遥感影像 + 分类　；参考：《新疆大学》2017年硕士论文

【摘要】：高精度的遥感影像分类是遥感技术应用于其他领域的重要前提。伴随着遥感技术的飞速发展,利用遥感技术对地物信息进行分类一直是遥感领域的研究热点。近年来,越来越多的学者将机器学习算法应用于模式识别领域。其中,浅层机器学习算法在遥感影像分类方面取得了一定的成果,但是由于浅层机器学习的学习能力有限且泛化能力比较差,其效率并不能很好的满足海量遥感影像分类的需求。深度学习的兴起及迅速发展,为海量遥感影像分类的研究开辟了新的道路。分布式计算框架的提出为遥感影像分类的高效并行提供了有效的解决方案。本文从浅层机器学习、深度学习和并行计算三个方面对遥感影像分类进行研究,主要研究内容有以下三方面:(1)构建了一种基于BP神经网络的遥感影像棉花识别模型。为了简化使用遥感影像进行棉花识别的过程,提高棉花识别的精度,本文提出了基于BP神经网络的棉花识别方法。从高分一号8米分辨率影像中提取归一化植被指数、差值植被指数、比值植被指数等多个特征指数,并将不同的特征指数进行组合作为BP神经网络的输入,对网络进行训练。实验表明:该模型能有效提高遥感影像棉花识别的精度。(2)构建了一种基于卷积神经网络的遥感影像草地分类模型。针对浅层机器学习在处理海量遥感数据时存在的局限性,本文将深度学习思想引入遥感影像分类的研究之中,提出了基于卷积神经网络的遥感影像分类模型。该模型可以直接将遥感图像作为网络的输入,通过逐层的网络训练,提取有用的特征用以实现遥感影像的自动分类。实验结果表明:迭代次数的增加及训练样本的增加均可提高模型的分类精度;基于卷积神经网络的遥感影像草地分类模型的分类精度为96.35%,比支持向量机(SVM)和BP神经网络方法分别提高8.90%和4.78%。(3)基于Spark分布式内存计算平台实现了遥感影像分类的并行化。针对卷积神经网络模型在进行遥感影像草地分类过程中的高时耗问题,本文将并行计算思想引入到遥感影像分类的研究之中。借助Hadoop分布式文件系统(HDFS)对遥感影像进行分布式存储,通过Spark分布式内存计算平台实现遥感影像分类的高效并行。在分布式计算集群上采用数据并行方式实现卷积神经网络的并行化模型。通过三组实验验证了并行化卷积神经网络模型的性能。实验结果表明:并行化的卷积神经网络模型在保证遥感影像草地分类精度的同时,还提高了草地分类的效率,实现了真正意义上的高效并行计算。
[Abstract]:High precision classification of remote sensing images is an important prerequisite for the application of remote sensing technology in other fields. With the rapid development of remote sensing technology, the classification of ground object information using remote sensing technology has been a hot spot in the field of remote sensing. In recent years, more and more scholars apply machine learning algorithm to pattern recognition. Among them, the shallow machine learning algorithm has made some achievements in remote sensing image classification, but the learning ability of shallow machine learning is limited and the generalization ability is relatively poor. Its efficiency can not meet the needs of massive remote sensing image classification. The rise and rapid development of deep learning has opened a new way for the classification of massive remote sensing images. The proposed distributed computing framework provides an effective solution for efficient parallel classification of remote sensing images. In this paper, the classification of remote sensing images is studied from three aspects: shallow machine learning, deep learning and parallel computing. The main research contents are as follows: 1) A cotton recognition model based on BP neural network is constructed. In order to simplify the process of cotton recognition using remote sensing image and improve the precision of cotton recognition, a cotton recognition method based on BP neural network is proposed in this paper. The normalized vegetation index, difference vegetation index and ratio vegetation index were extracted from the 8m resolution image of Gao Fen 1, and the different characteristic indexes were combined as the input of BP neural network to train the network. Experiments show that the model can effectively improve the precision of cotton recognition in remote sensing images. (2) A remote sensing image grassland classification model based on convolution neural network is constructed. Aiming at the limitation of shallow machine learning in processing massive remote sensing data, this paper introduces the idea of deep learning into the research of remote sensing image classification, and puts forward a classification model of remote sensing image based on convolution neural network. The model can directly take remote sensing images as the input of the network and extract useful features to realize the automatic classification of remote sensing images through the training of the network layer by layer. The experimental results show that the classification accuracy of the model can be improved by increasing the number of iterations and training samples. The classification accuracy of remote sensing image grassland classification model based on convolution neural network is 96.35, which is 8.90% and 4.78% higher than support vector machine (SVM) and BP neural network, respectively. The parallel classification of remote sensing image is realized based on Spark distributed memory computing platform. Aiming at the problem of high time consumption in remote sensing image grassland classification based on convolution neural network model, this paper introduces the idea of parallel computing into the research of remote sensing image classification. Remote sensing images are stored distributed by Hadoop distributed file system (HDFS), and the efficient parallel classification of remote sensing images is realized by Spark distributed memory computing platform. The parallel model of convolutional neural network is realized by data parallelism in distributed computing cluster. The performance of parallel convolution neural network model is verified by three groups of experiments. The experimental results show that the parallel convolution neural network model not only ensures the accuracy of grassland classification in remote sensing images, but also improves the efficiency of grassland classification.
【学位授予单位】：新疆大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】