深度卷积神经网络特征提取用于地表覆盖分类初探

发布时间：2018-03-20 09:01

本文选题：卷积神经网络　切入点：AlexNet　出处：《中国图象图形学报》2017年08期 　论文类型：期刊论文

【摘要】：目的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,在全球资源监测、全球变化检测中发挥着重要作用。提高中等分辨率遥感影像地表覆盖分类的精度具有非常重要的意义。方法近年来,深度卷积神经网络在图像分类、目标检测和图像语义分割等领域取得了一系列突破性的进展,相比于传统的机器学习方法具有更强的特征学习和特征表达能力。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。以GF-1的16 m空间分辨率多光谱影像为实验数据,利用预训练好的Alex Net深度卷积神经网络模型进行特征提取,以SVM为分类器进行分类。分析了Alex Net不同层的特征以及用于提取特征的邻域窗口尺寸对分类结果的影响,并与传统的单纯基于光谱特征和基于光谱+纹理特征的分类结果进行对比分析。结果结果表明在用Alex Net模型提取特征进行地表覆盖分类时,Fc6全连接层是最有效的特征提取层,最佳的特征提取窗口尺寸为9×9像素,同时利用深度特征得到的总体分类精度要高于其他两种方法。结论深度卷积神经网络可以提取更精细更准确的地表覆盖特征,得到更高的地表覆盖分类精度,为地表覆盖分类提供了参考价值。
[Abstract]:Destination table coverage monitoring is an important basis for eco-environmental change research, land resource management and sustainable development, in the context of global resource monitoring, Global change detection plays an important role. It is of great significance to improve the accuracy of ground cover classification of middle resolution remote sensing images. Methods in recent years, deep convolution neural networks are used in image classification. A series of breakthrough advances have been made in the fields of target detection and image semantic segmentation. Compared with the traditional machine learning method, it has stronger feature learning and feature expression ability. In this paper, the feature extraction and classification of middle resolution remote sensing images based on deep convolution neural network are studied. The 16 m spatial resolution multispectral images of GF-1 are used as experimental data. Using the pre-trained Alex Net deep convolution neural network model for feature extraction and SVM as the classifier, the features of different layers of Alex Net and the influence of the size of the neighborhood window used to extract the features on the classification results are analyzed. The results are compared with the traditional classification results based on spectral features and spectral texture features. The results show that Fc6 fully connected layer is the most effective feature extraction layer when using Alex Net model to extract features for surface cover classification. The optimal feature extraction window size is 9 脳 9 pixels, and the overall classification accuracy obtained by using depth features is higher than the other two methods. Conclusion Deep convolution neural network can extract more precise and accurate surface cover features. Higher accuracy of surface cover classification is obtained, which provides reference value for surface cover classification.
【作者单位】：中国科学院遥感与数字地球研究所;中国科学院大学;
【基金】：高分辨率对地观测系统重大专项(03-Y20A04-9001-15/16,11-Y20A05-9001-15/16)~~
【分类号】：TP183;TP751

【相似文献】