关键词: 遥感影像分类, 深度神经网络, 迁移学习, VGG16, SegNet, 传递迁移学习, 资源三号影像, 全局平均池化

Abstract:

In the practical task of remote sensing image classification, the biggest problem with the use of deep neural network method is the lack of sufficient labeled samples. How to use fewer labeled samples to achieve higher accuracy of remote sensing image classification is a problem that needs to be solved at present. ImageNet is the largest image recognition dataset in the world, the model trained on it has rich underlying features. Fine-tuning the ImageNet pre-training model is the most common transfer learning method, which can make use of the rich underlying features to improve the classification accuracy. However, there is a big difference between ImageNet image features and remote sensing image features, and the improvement of classification effect is limited. In order to solve the above problems, a remote sensing image classification method based on deep transitive transfer learning combined with deep neural network is proposed in this paper. This method constructs an intermediate domain using the open-source remote sensing scene recognition datasets as the data source and uses ImageNet pre-training weight as the source domain and remote sensing images to be classified as the target domain for transfer learning to improve remote sensing image classification accuracy. First, based on ImageNet pre-training VGG16 network, the fully connected layer is replaced by the global average pooling layer in order to speed up the weight update of convolutional layer, and the GAP-VGG16 is constructed. The intermediate domain dataset is used for training the ImageNet pre-training GAP-VGG16 to obtain the weight. Then, based on the SegNet, the T-SegNet is designed by adding the convolutional layer into the SegNet to further extract the obtained weight. Finally, the obtained weight is transferred to T-SegNet, and the remote sensing image classification is achieved by training the target domain dataset. In this paper, the Aerial Image Dataset and UC Merced Land-Use Dataset are selected as the data sources of the intermediate domain dataset, and the ZY-3 Panjin area image is selected as the target domain image, 50% and 25% of the training samples are selected for the experiment. The experimental results show that using 50% and 25% of the training samples, the Kappa coefficient of the classification results using the proposed method in this paper is increased by 0.0459 and 0.0545, respectively compared to SegNet, and is increased by 0.0377 and 0.0346, respectively compared to ImageNet pre-training SegNet. For classes with a smaller number of samples, the classification accuracy of the method in this paper is improved more significantly.

Key words: remote sensing classification, deep neural network, transfer learning, VGG16, SegNet, transitive transfer learning, ZY-3 remote sensing image, global average pooling