高分辨卫星图像卷积神经网络分类模型

发布时间：2018-08-01 13:01

【摘要】：目的卫星图像往往目标、背景复杂而且带有噪声,因此使用人工选取的特征进行卫星图像的分类就变得十分困难。提出一种新的使用卷积神经网络进行卫星图像分类的方案。使用卷积神经网络可以提取卫星图像的高层特征,进而提高卫星图像分类的识别率。方法首先,提出一个包含六类图像的新的卫星图像数据集来解决卷积神经网络的有标签训练样本不足的问题。其次,使用了一种直接训练卷积神经网络模型和3种预训练卷积神经网络模型来进行卫星图像分类。直接训练模型直接在文章提出的数据集上进行训练,预训练模型先在ILSVRC(the Image Net large scale visual recognition challenge)-2012数据集上进行预训练,然后在提出的卫星图像数据集上进行微调训练。完成微调的模型用于卫星图像分类。结果提出的微调预训练卷积神经网络深层模型具有最高的分类正确率。在提出的数据集上,深层卷积神经网络模型达到了99.50%的识别率。在数据集UC Merced Land Use上,深层卷积神经网络模型达到了96.44%的识别率。结论本文提出的数据集具有一般性和代表性,使用的深层卷积神经网络模型具有很强的特征提取能力和分类能力,且是一种端到端的分类模型,不需要堆叠其他模型或分类器。在高分辨卫星图像的分类上,本文模型和对比模型相比取得了更有说服力的结果。
[Abstract]:Objective Satellite images often have complex background and noise, so it is very difficult to classify satellite images by using artificially selected features. A new scheme of satellite image classification using convolution neural network is proposed. Convolution neural network can be used to extract the high-level features of satellite images, and thus improve the recognition rate of satellite image classification. Firstly, a new satellite image data set including six kinds of images is proposed to solve the problem of the shortage of labeled training samples in convolutional neural networks. Secondly, a direct training convolution neural network model and three pre-training convolution neural network models are used to classify satellite images. The direct training model is trained directly on the data set proposed in this paper. The pre-training model is pretrained on the ILSVRC (the Image Net large scale visual recognition challenge) -2012 data set and then fine-tuned on the proposed satellite image data set. The fine tuning model is used for satellite image classification. Results the proposed model has the highest classification accuracy. In the proposed data set, the recognition rate of the deep convolution neural network model is 99.50%. On the data set UC Merced Land Use, the recognition rate of the deep convolution neural network model is 96.44%. Conclusion the data set presented in this paper is general and representative. The deep convolution neural network model has strong ability of feature extraction and classification. It is an end-to-end classification model and does not need to stack other models or classifiers. In the classification of high resolution satellite images, the proposed model is more persuasive than the contrast model.
【作者单位】：陕西师范大学计算机科学学院;
【基金】：国家自然科学基金项目(41171338,41471280,61401265)~~
【分类号】：TP183;TP391.41

【相似文献】