Stochastic and confidence-aware network (SCAN)-based semisupervised domain adaptation for satellite imagery...

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Satellite imagery segmentation is vital for environmental monitoring but faces challenges due to domain gaps and high annotation costs.
• Existing semisupervised domain adaptation methods can overfit limited labeled target data.

Purpose of the Study:

• To propose a novel confidence-aware semisupervised domain adaptation method for satellite imagery segmentation.
• To address the limitations of domain gap and high annotation costs in practical satellite image analysis.

Main Methods:

• A confidence-aware semisupervised domain adaptation approach that refines pseudolabel confidence using labeled target data.
• Incorporation of a sparsity constraint to select robust and compact features generalizable across domains.
• Leveraging a teacher model to generate and refine pseudolabels, mitigating overfitting on small labeled datasets.

Main Results:

• The method achieves 98.04% and 93.78% of fully supervised performance on LoveDA and Open Earth Map datasets using only 5% labeled target data.
• Evaluated on LoveDA, SyntheWorld, and Open Earth Map datasets, demonstrating robust performance across different scenarios.
• Qualitative experiments confirm the method avoids overfitting and produces reasonable segmentation results.

Conclusions:

• The proposed confidence-aware semisupervised domain adaptation method effectively reduces annotation requirements for satellite imagery segmentation.
• The approach demonstrates strong generalization capabilities and robust performance in the presence of domain shifts.
• This work offers a promising solution for practical environmental monitoring applications with limited labeled data.