S³CD: A Self-Supervised Semantic Change Detection Method by Mining Transition Patterns and Consistency in Remote...

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Semantic change detection (SCD) is crucial for monitoring land-cover changes using remote sensing images.
• Traditional supervised SCD methods require extensive pixel-level annotations, hindering their practical application.
• Self-supervised learning (SSL) offers a promising alternative by leveraging large unlabeled datasets, but existing SSL methods struggle with complex semantic transformations in SCD.

Purpose of the Study:

• To develop an effective self-supervised network for semantic change detection (SCD) in bi-temporal remote sensing images.
• To address the limitations of existing SSL methods in capturing complex semantic transformations required for SCD.
• To reduce the reliance on pixel-level annotations in remote sensing change detection tasks.

Main Methods:

• Propose a multistage, multitask, and multilevel self-supervised network (S3CD) for SCD.
• The network learns semantic changes across scene, pixel, and prototype levels in two stages.
• Stage 2 enhances robustness by learning semantic consistency and capturing temporal patterns of change categories.

Main Results:

• S3CD demonstrated superior performance compared to state-of-the-art self-supervised and supervised SCD methods on two benchmark datasets.
• The proposed method achieved competitive results in binary change detection (BCD), outperforming most supervised learning approaches.
• S3CD effectively learns semantic representations for complex land-cover changes without extensive manual annotation.

Conclusions:

• The developed S3CD network significantly advances self-supervised learning for remote sensing change detection.
• This approach offers a viable solution for large-scale land-cover monitoring by minimizing annotation requirements.
• S3CD facilitates the broader application of self-supervised methods in geospatial analysis and change detection.