Cross-Domain Land Cover Classification of Remote Sensing Images Based on Full-Level Domain Adaptation

The use of remote sensing images for land cover classification is an important and challenging pixel-level classification task. However, the different distribution of the same land cover categories across different datasets, the accuracy of the classification is significantly reduced when a classification model trained on one dataset is used directly on another dataset. To address the issue, numerous unsupervised domain adaptation (UDA) methods have been proposed. However, the existing UDA methods focus mainly on natural images and are not suited to remote sensing images with large variations in spectral information and texture features. Therefore, we develop a new full-level domain adaptation network (FLDA-NET) applicable for cross-domain land cover classification of remote sensing images. It aligns the source domain and target domain through a two-stage process encompassing image-level, feature-level, and output-level. In stage I, we align at image-level by converting the source domain image to the target domain image style. In stage II, at the feature-level we align the entropy of the two domain features. At the output-level we do not use simple global alignment, but category-level alignment. Furthermore, a self-training strategy based on superpixel segmentation and softmax probability is proposed to further enhance the model's performance on the target domain. Extensive experiments on our proposed FLDA-NET are performed on the Potsdam and Vaihingen datasets and compared with other advanced UDA methods. The outcome demonstrates that this approach greatly improves the ability of cross-domain land cover classification in remote sensing images.