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

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Cross-domain segmentation in remote sensing is challenged by variations in spectra, scale, and resolution.
• Existing domain adaptation methods often prioritize style over semantic information or overlook remote sensing specifics, leading to unstable training and biased results.

Purpose of the Study:

• To propose a novel dual-space generative adversarial domain adaptation segmentation framework, DS-DWTGAN, to minimize domain differences.
• To improve the performance and stability of semantic segmentation models across different remote sensing domains.

Main Methods:

• Introduced a dual-space generative adversarial network incorporating a wavelet transform branch to capture frequency and semantic information.
• Integrated output adaptation and data enhancement strategies for domain-invariant feature acquisition.
• Leveraged wavelet transform to preserve semantic details in the frequency domain, mitigating conversion deviations.

Main Results:

• Achieved superior performance on the PotsdamIRRG to Vaihingen task, with mIoU of 56.04% and mF1 of 67.28%.
• Outperformed state-of-the-art methods by 2.81% (mIoU) and 2.08% (mF1).
• Demonstrated enhanced model stability and reduced noise interference during domain migration.

Conclusions:

• The proposed DS-DWTGAN framework effectively addresses challenges in cross-domain semantic segmentation for remote sensing images.
• The wavelet transform integration and adaptation strategies significantly improve model adaptability and performance.
• The method shows superior efficacy for unsupervised semantic segmentation of UAV remote sensing images compared to existing approaches.