Unsupervised domain adaptation for remote sensing semantic segmentation with the 2D discrete wavelet transform

Abstract

Abstract There would be the differences in spectra, scale and resolution between the Remote Sensing datasets of the source and target domains, which would lead to the degradation of the cross-domain segmentation performance of the model. Image transfer faced two problems in the process of domain-adaptive learning: overly focusing on style features while ignoring semantic information, leading to biased transformation results, and easily overlooking the true transfer characteristics of remote sensing images, resulting in unstable model training. To address these issues, we proposes a novel dual-space generative adversarial domain adaptation segmentation framework, DS-DWTGAN, to minimize the differences between the source domain and the target domain. DS-DWTGAN aims to mitigate the distinctions between the source and target domains, thereby rectifying the imbalances in style and semantic representation.The framework introduces a network branch leveraging wavelet transform to capture comprehensive frequency domain and semantic information. It aims to preserve semantic details within the frequency domain space, mitigating image conversion deviations. Furthermore, our proposed method integrates output adaptation and data enhancement training strategies to reinforce the acquisition of domain-invariant features. This approach effectively diminishes noise interference during the migration process, bolsters model stability, and elevates the model’s adaptability to remote sensing images within different domains. Experimental validation was conducted on the publicly available Potsdam and Vaihingen datasets. The findings reveal that in the PotsdamIRRG to Vaihingen task, the proposed method attains outstanding performance with mIoU and mF1 values reaching 56.04% and 67.28%, respectively. Notably, these metrics surpass the corresponding values achieved by state-of-the-art (SOTA) methods, registering an increase of 2.81% and 2.08%. In comparison to alternative approaches, our proposed framework exhibits superior efficacy in the domain of unsupervised semantic segmentation for UAV remote sensing images.