High spatial resolution imagery scene classification based on semi-supervised CNNs

The large amount of labeled dataset is always required to train the deep convolutional neural networks (CNNs) for high spatial resolution (HSR) imagery scene classification. However, the generalization of the learned deep features is decayed when limited labeled data is available. To solve this problem, the scene classification based on semi-supervised CNNs (3sCNN) is proposed. In the proposed method, the labeled data is first used to train the model and then the prediction label and confidence of the unlabeled data is obtained with the trained model. Finally, the unlabeled data with high confidence is considered as the labeled data to train the network again and the progress described above is repeated. To demonstrate the effectiveness of the proposed method, the experiments on three datasets are performed. The results show that the proposed method can effectively improve the classification.