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。文章提出了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，提出了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三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 ImageNet Large Scale Visual Recognition Challenge)-2012数据集上进行预训练，然后在文章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：文章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。在文章提出的数据集上，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：文章提出的数据集具有一般性和代表性，文章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，文章提出的模型和对比模型相比取得了更有说服力的结果。
Convolutional neural network models for high spatial resolution satellite imagery classification
Zhou Mingfei,Wang Xili,Wang Lei,Chen Fen(School of Computer Science,Shaanxi Normal University,Xi’an City,Shaanxi Province)
Objective: Satellite imagery classification is the task that use classification models to divide a set of satellite images into several classes. The satellite images discussed in this paper are collected from the Quickbird satellite imagery dataset. In this paper, satellite images are divided into six classes, which are airplanes, dense residential areas, harbors, intersections, overpasses and parking lots. Generally, the task of satellite imagery classification is difficult due to the complex targets and backgrounds in satellite images. Traditional methods, such as artificial neural networks and support vector machines, usually use low level features and chosen features by manual. These features are insufficient and they cannot represent the multi-level and intrinsic features of satellite images. At the same time, classification methods which use these low level features are difficult to obtain high accuracy. Some deep learning methods use pre-trained convolutional neural networks to extract the high-level features of satellite images and some classifier to classify satellite images. They can improve performance than traditional methods. But these methods ignore the inherent classification ability of convolutional neural networks. This is because a large amount of labeled training data of satellite images is needed to train a convolutional neural network which could both extract features and classify images simultaneously, but the training data is limited in practice. Some other methods use a stack of shallow convolutional neural networks to classify satellite images. But the stack of low level features is still not representative enough to improve the classification accuracy of satellite image substantially. In this paper, a new approach using deep convolutional neural networks is presented to improve the classification accuracy for satellite imagery. The classification accuracy of satellite images could be improved using the deep features extracted by convolutional neural networks. Method: An end-to-end training and classification method is proposed in this paper. Additional classifier is not needed in this method. Meanwhile, this method does not need the stack of shallow convolutional neural networks to improve the ability of feature extraction from satellite images. Firstly, a new satellite imagery dataset which contains six classes is proposed to deal with the problem of lacking labeled training data in this paper. Secondly, three kinds of pre-trained deep convolutional neural network models and a directly trained shallow convolutional neural network model are used to perform the classification task for satellite images. The shallow model has a small amount of training weights and can be trained directly on the satellite image dataset which is proposed in this paper to classify satellite images. The proposed three kinds of deep models in this paper should be pre-trained on an auxiliary dataset. This is because the amount of training weights of the three deep models is too large to be directly trained on the proposed satellite images dataset. The three kinds of deep models are pre-trained on a large auxiliary dataset which contains roughly one million and two hundred thousand labeled training images of one thousand classes. All of the images contain the common objects which could be seen everywhere in the daily life. The weights of the three deep architectures of convolutional neural networks can be trained adequately after pre-training on the large auxiliary dataset. The ability of the deep models to extract representative features and to classify images can be improved after pre-training, and the application objects of the models can be transferred from daily common objects to satellite image objects. The key point of such transformation is fine-tuning the pre-trained deep models on the proposed satellite images dataset. The architectures of the three deep models should be changed slightly and then they could be fine-tuned on the proposed dataset. After fine-tuning, the three deep convolutional neural network models could be used to classify satellite images directly without the help of other classifiers or stacked shallow models. Result: The proposed convolutional neural network models are validated on two datasets. One of the testing satellite images dataset is the proposed dataset in this paper, the other dataset is the famous UC Merced Land Use dataset. All four proposed models obtain high performance on the proposed dataset in this paper. The classification accuracies of three deep models are higher than the accuracy of the shallow model. In particular, the deepest convolutional neural network model achieves the highest accuracy of 99.50% on the proposed dataset. The results on UC Merced Land Use dataset of three similar methods in literatures are compared with the results of the proposed models in this paper. Two of the three comparative methods use the features extracted from pre-trained convolutional neural networks without fine-tuning and use additional classifier to classify satellite images. The other one method uses a multi-view convolutional neural network to perform the classification task. Experimental results show that the proposed deep models in this paper achieve the highest accuracy (96.44%) among all the models. Conclusion: In this paper, a new satellite images dataset is proposed which is representative of satellite images. Convolutional neural networks could be trained adequately with the help of the proposed satellite image dataset. It is possible for shallow convolutional neural networks to be directly trained on this dataset. Pre-trained convolutional neural networks obtain better classification accuracy on other satellite imagery dataset after fine-tuning on the proposed dataset. Also, the proposed deep convolutional neural network models are effective in terms of deep feature extraction and satellite images classification. The proposed models obtain more competitive results comparing with other methods in literatures mentioned in this paper. The proposed deep models have good generalization ability and could achieve high accuracy on the UC Merced Land Use dataset whose images are different with the fine-tuning dataset in scale and quality. The effective pre-training and fine-tuning together with the depth of the proposed deep models in this paper contribute to the good performance. In addition, the proposed models are end-to-end models. Extra classifiers and the stack of shallow models are not needed to classify satellite images.