基于深度卷积神经网络自学习特征的地表覆盖分类研究

发布时间：2018-06-07 01:33

本文选题：深度卷积神经网络 + 特征提取　；参考：《中国科学院大学(中国科学院遥感与数字地球研究所)》2017年硕士论文

【摘要】：地表覆盖监测是生态环境变化研究、土地资源管理和可持续发展的重要基础。它能真实地反映地表覆盖情况,与人类的生产和生活息息相关。利用遥感手段获得地表覆盖信息的一个重要方法就是分类。传统的分类方法仅仅利用像元的光谱特征进行分类,不容易解决同物异谱、异物同谱以及混合像元等问题,且分类结果中的图斑比较零乱,常常出现错分、漏分的现象,从而造成分类精度不高。在光谱特征的基础上加入纹理特征进行分类,精度有所提高,但对于中低分辨率影像的地表覆盖分类,纹理特征对分类精度的提高作用有限,分类结果很难达到满意的效果。因此如何提取出更具表达性的特征是遥感影像分类的关键,选择与提取出的特征表达性越强,则分类效果会更好。近年来,深度卷积神经网络在图像分类、人脸识别、目标检测和图像语义分割等领域取得了巨大成果。它的局部感受野、权值共享和池化操作等特性可以有效地减少模型的训练参数,降低网络的复杂度,使模型对平移、扭曲和缩放具有一定程度的不变性。与传统机器学习方法相比具有更强大的特征学习和特征表达能力,以及更强的强鲁棒性和容错能力。基于其优越的特性,本文进行了深度卷积神经网络对中分辨率遥感影像进行特征提取和分类的探索性研究,提出了一种基于深度卷积神经网络自学习特征的地表覆盖分类方法。以支持向量机为分类器,在GF-1的WFV16米空间分辨率多光谱影像上进行分类实验,分析了不同深度卷积神经网络模型、深度卷积神经网络不同层的特征以及提取特征的邻域窗口尺寸对分类结果的影响。并与传统的仅基于光谱特征和基于光谱+纹理特征的分类结果进行了对比与分析,证明了从深度卷积神经网络提取的深度特征更具表达性,能更准确地反映地表覆盖的特征,得到更高的分类精度,为地表覆盖分类提供了参考价值。
[Abstract]:Surface cover monitoring is an important basis for the study of ecological environment change, land resource management and sustainable development. It can truly reflect the surface cover, and is closely related to human production and life. Classification is an important method for obtaining land cover information by remote sensing. The traditional classification methods only use the spectral characteristics of pixels to classify. It is not easy to solve the problems of isospectral, foreign body isospectral and mixed pixel. As a result, the classification accuracy is not high. On the basis of spectral features, the classification accuracy is improved by adding texture features. However, for the ground cover classification of low and medium resolution images, the effect of texture features on classification accuracy is limited, and the classification results are difficult to achieve satisfactory results. Therefore, how to extract more expressive features is the key to the classification of remote sensing images. The stronger the expression of the selected and extracted features, the better the classification effect will be. In recent years, deep convolution neural networks have made great achievements in image classification, face recognition, target detection and image semantic segmentation. Its local receptive field, weight sharing and pool operation can effectively reduce the training parameters of the model, reduce the complexity of the network, and make the model invariant to translation, distortion and scaling to a certain extent. Compared with traditional machine learning methods, it has more powerful feature learning and feature representation ability, stronger robustness and fault tolerance. Based on its superior characteristics, this paper makes an exploratory study on feature extraction and classification of middle resolution remote sensing images based on deep convolution neural network, and proposes a surface cover classification method based on self-learning features of deep convolution neural network. Using support vector machine (SVM) as classifier, the neural network model of convolution in different depth is analyzed on the WFV16 meter spatial resolution multispectral image of GF-1. The influence of the features of different layers of the deep convolution neural network and the size of the neighborhood window of the feature extraction on the classification results. Compared with the traditional classification results based on spectral features and spectral texture features, it is proved that the depth features extracted from the deep convolution neural network are more expressive and can reflect the features of surface cover more accurately. Higher classification accuracy is obtained, which provides reference value for surface cover classification.
【学位授予单位】：中国科学院大学(中国科学院遥感与数字地球研究所)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751;TP183

【相似文献】