Cross-Domain Few-Shot Segmentation for Remote Sensing Image Based on Task Augmentation and Feature Disentanglement
Jiehu Chen, Xili Wang, Ling En Hong, Ming Liu · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2024
Few-shot segmentation aims to segment a large number of unlabeled samples in the target domain, by leveraging the images and labels from the source domain as well as a few labeled samples from the target domain. This is pivotal in tackling the scarcity of labeled samples in remote sensing image segmentation tasks. However, prevalent few-shot segmentation methods overlook inter-domain discrepancies, do not model and leverage the relationship between samples, and often only implement binary classification but not multi-class classification directly. To address these problems, we propose a cross-domain few-shot segmentation method based on task augmentation and feature disentanglement for practical remote sensing segmentation tasks. On one hand, task augmentation, which involves increasing the diversity of the training set and generating more challenging training data, can improve the model's generalization. On the other hand, feature disentanglement, involving the extraction of domain-irrelevant features for segmentation, improves the transferability of the model. Furthermore, to flexibly capture the relationships between the segmented regions, a graph with regions as nodes and relationships between nodes as edges is constructed. Then, labels are propagated from the labeled nodes to the unlabeled nodes in the graph by label propagation algorithm to implement multi-class classification directly. We conducted experiments on two public datasets, WHDLD and GID5 as well as a Tibetan Plateau dataset collected by our group, by the proposed method and several comparative methods. The average of three experimental results show that the proposed method outperforms the suboptimal results by 5.99%, 6.12%, and 11.34% in terms of mIoU, mean F1, and OA when the labeled samples are 5, and by 8.97%, 9.04%, and 14.29% in mIoU, mean F1, and OA for 10 labeled samples.