SOURCE-FREE DOMAIN ADAPTATION FOR REMOTE SENSING IMAGE SEGMENTATION VIA TARGET DOMAIN DATA ENHANCEMENT

Remote sensing image segmentation is a vital but challenging task. The high cost of expert annotations and the limited availability of labeled data significantly hinder the training process. The Unsupervised Domain Adaptation method can solve the problem of unlabeled target domain image segmentation in the case of domain shift. However, concerns regarding national security and privacy restrict access to source domain data. The task of performing high-quality segmentation on unlabeled target data using a source domain model is challenging. To address these issues, we propose a source-free domain adaptation method for remote sensing image segmentation using target domain data enhancement. We improve the segmentation performance of the target domain model by augmenting the target domain pseudo labels and paired data. For pseudo-label denoising, we use binary pseudo labels to obtain class prototypes and assess the reliability of the pseudo labels by measuring the relative feature distance to the class prototypes. In the adaptation stage, we enhance the robustness of the model by rotating both the target domain data and pseudo labels. This method was tested on a pair of remote sensing datasets: Potsdam IR-R-G to Vaihingen IR-R-G. The results demonstrate significant improvements, with the method achieving a 13.89% increase in MioU, a 19.11% improvement in OA, and a 19.81% increase in the kappa coefficient, compared to direct segmentation using the source domain training model. These experimental results highlight the effectiveness of the proposed method in the source-free domain adaptation scenario.