Remote sensing image scene classification takes image blocks as classification units and predicts their semantic descriptors. Because it is difficult to obtain enough labeled samples for all classes of remote sensing image scenes, zero-shot classification methods which can recognize image scenes that are not seen in the training stage are of great significance. By projecting the image visual features and the class semantic features into the latent space and ensuring their alignment, the variational autoencoder (VAE) generative model has been applied to address remote-sensing image scene classification under a zero-shot setting. However, the VAE model takes the element-wise square error as the reconstruction loss, which may not be suitable for measuring the reconstruction quality of the visual and semantic features. Therefore, this paper proposes to augment the VAE models with the generative adversarial network (GAN) to make use of the GAN’s discriminator in order to learn a suitable reconstruction quality metric for VAE. To promote feature alignment in the latent space, we have also proposed cross-modal feature-matching loss to make sure that the visual features of one class are aligned with the semantic features of the class and not those of other classes. Based on a public dataset, our experiments have shown the effects of the proposed improvements. Moreover, taking the ResNet models of ResNet18, extracting 512-dimensional visual features, and ResNet50 and ResNet101, both extracting 2048-dimensional visual features for testing, the impact of the different visual feature extractors has also been investigated. The experimental results show that better performance is achieved by ResNet18. This indicates that more layers of the extractors and larger dimensions of the extracted features may not contribute to the image scene classification under a zero-shot setting.

Zero-shot classification of image scenes which can recognize the image scenes that are not seen in the training stage holds great promise of lowering the dependence on large numbers of labeled samples. To address the zero-shot image scene classification, the cross-modal feature alignment methods have been proposed in recent years. These methods mainly focus on matching the visual features of each image scene with their corresponding semantic descriptors in the latent space. Less attention has been paid to the contrastive relationships between different image scenes and different semantic descriptors. In light of the challenge of large intra-class difference and inter-class similarity among image scenes and the potential noisy samples, these methods are susceptible to the influence of the instances which are far from these of the same classes and close to these of other classes. In this work, we propose a multi-level cross-modal feature alignment method via contrastive learning for zero-shot classification of remote sensing image scenes. While promoting the single-instance level positive alignment between each image scene with their corresponding semantic descriptors, the proposed method takes the cross-instance contrastive relationships into consideration,and learns to keep the visual and semantic features of different classes in the latent space apart from each other. Extensive experiments have been done to evaluate the performance of the proposed method. The results show that our proposed method outperforms state of the art methods for zero-shot remote sensing image scene classification. All the code and data are available at github https://github.com/masuqiang/MCFA-Pytorch