Semantic Segmentation of Remote Sensing Images via Visible-Infrared Domain Adaptation

Segmentation of infrared images is frequently essential in remote sensing image processing. However, infrared imagery poses significant challenges compared to conventional visible image segmentation, low resolution, insufficient data and poor labeling quality. The unsupervised domain adaptation method becomes an effective solution to this problem, which aims to enable a model trained on a source domain dataset with labeled information to adapt to another unlabeled target dataset. In this paper, the research of unsupervised domain adaptation method for the semantic segmentation of remote sensing images is carried out based on deep learning. The training process of a deep convolutional network requires a large amount of labeled data. However, remote sensing images are affected by differences in data acquisition location, time, light angle and other optical influences. To address this challenge, this paper proposes a remote sensing image domain adaptation semantic segmentation network that combines surface features and deep features. Photometric alignment is employed to intuitively reduce the domain gap and align surface features, while prototype-based classification and contrastive learning are used to align deep features. The experimental results demonstrate that this method achieves effective and reliable segmentation results on remote sensing image data from different scenes and improves the generalization ability of the semantic segmentation model.