DFEN: A Dual-Feature Extraction Network-Based Open-Set Domain Adaptation Method for Optical Remote Sensing Image Scene Classification

Open-set domain adaptation methods aim to correctly classify data when there is a disparity in distribution and class spaces between the test and training sets. However, existing methods mainly concentrate on natural images, and their performance is hindered in more complex remote sensing image scene classification tasks by various inherent factors, in particular, diverse imaging conditions, different resolutions, and geographical semantics. To address this issue, we propose a dual-feature extraction network-based open-set domain adaptation method (DFEN). Specifically, we design a dual-feature extraction network model, comprising a local feature extractor and a global feature extractor. The local feature extractor is primarily used to capture local distinctive features from images to enhance the discriminative ability of the model for similar categories and improve its resistance to interference in multi-object images. By contrast, the global feature extractor focuses on extracting semantic correlations that are independent of image style and target scale to improve the model's understanding and generalization ability for scene categories. Besides, to boost the classification accuracy for unknown categories in open environments, we introduce an adaptive unknown class weighting mechanism based on the similarity between samples and known classes. By comprehensively measuring across the entire classification space and individual categories, samples with high weights are considered as unknown class. Experimental results demonstrate that the proposed method achieves promising performance in various open and cross-domain scenarios.