Remote sensing image scene classification is one of the most challenging problems in understanding high-resolution remote sensing images. Deep learning techniques, especially the convolutional neural network (CNN), have improved the performance of remote sensing image scene classification due to the powerful perspective of feature learning and reasoning. However, several fully connected layers are always added to the end of CNN models, which is not efficient in capturing the hierarchical structure of the entities in the images and does not fully consider the spatial information that is important to classification. Fortunately, capsule network (CapsNet), which is a novel network architecture that uses a group of neurons as a capsule or vector to replace the neuron in the traditional neural network and can encode the properties and spatial information of features in an image to achieve equivariance, has become an active area in the classification field in the past two years. Motivated by this idea, this paper proposes an effective remote sensing image scene classification architecture named CNN-CapsNet to make full use of the merits of these two models: CNN and CapsNet. First, a CNN without fully connected layers is used as an initial feature maps extractor. In detail, a pretrained deep CNN model that was fully trained on the ImageNet dataset is selected as a feature extractor in this paper. Then, the initial feature maps are fed into a newly designed CapsNet to obtain the final classification result. The proposed architecture is extensively evaluated on three public challenging benchmark remote sensing image datasets: the UC Merced Land-Use dataset with 21 scene categories, AID dataset with 30 scene categories, and the NWPU-RESISC45 dataset with 45 challenging scene categories. The experimental results demonstrate that the proposed method can lead to a competitive classification performance compared with the state-of-the-art methods.
Land cover classification is one of the most fundamental tasks in the field of remote sensing. In recent years, fully supervised fully convolutional network (FCN)-based semantic segmentation models have achieved state-of-the-art performance in the semantic segmentation task. However, creating pixel-level annotations is prohibitively expensive and laborious, especially when dealing with remote sensing images. Weakly supervised learning methods from weakly labeled annotations can overcome this difficulty to some extent and achieve impressive segmentation results, but results are limited in accuracy. Inspired by point supervision and the traditional segmentation method of seeded region growing (SRG) algorithm, a weakly towards strongly (WTS) supervised learning framework is proposed in this study for remote sensing land cover classification to handle the absence of well-labeled and abundant pixel-level annotations when using segmentation models. In this framework, only several points with true class labels are required as the training set, which are much less expensive to acquire compared with pixel-level annotations through field survey or visual interpretation using high-resolution images. Firstly, they are used to train a Support Vector Machine (SVM) classifier. Once fully trained, the SVM is used to generate the initial seeded pixel-level training set, in which only the pixels with high confidence are assigned with class labels whereas others are unlabeled. They are used to weakly train the segmentation model. Then, the seeded region growing module and fully connected Conditional Random Fields (CRFs) are used to iteratively update the seeded pixel-level training set for progressively increasing pixel-level supervision of the segmentation model. Sentinel-2 remote sensing images are used to validate the proposed framework, and SVM is selected for comparison. In addition, FROM-GLC10 global land cover map is used as training reference to directly train the segmentation model. Experimental results show that the proposed framework outperforms other methods and can be highly recommended for land cover classification tasks when the pixel-level labeled datasets are insufficient by using segmentation models.