Unsupervised domain adaptive building semantic segmentation network by edge-enhanced contrastive learning.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Unsupervised domain adaptation (UDA) is crucial for image classification with limited labeled data.
• Extracting artificial features like buildings from high-resolution remote sensing (HSR) imagery using UDA remains challenging due to complex imaging conditions.

Purpose of the Study:

• To propose a novel UDA method, CDANet, for accurate building extraction in HSR imagery.
• To enhance building extraction by integrating adversarial and contrastive learning techniques.

Main Methods:

• CDANet employs a multitask generator with dual branches for region and edge extraction.
• Multilevel adversarial learning is achieved through dual discriminators processing prediction outputs.
• Regional pixelwise contrastive loss aligns cross-domain pixel features in the embedding space.
• A self-training strategy with pseudolabel generation addresses target domain discrepancies.

Main Results:

• Comprehensive experiments on WHU, Austin, and Massachusetts datasets validate CDANet's effectiveness.
• Ablation studies confirm improvements from the generator structure, contrastive loss, and self-training strategy.
• CDANet outperforms state-of-the-art methods like AdaptSegNet, AdvEnt, IntraDA, FDANet, and ADRS in F1 score and mIoU.

Conclusions:

• CDANet significantly improves building extraction accuracy in HSR imagery.
• The proposed method effectively addresses challenges in unsupervised domain adaptation for remote sensing.
• CDANet offers a robust solution for extracting small and densely distributed buildings.