SRDFNet: Semantic Refinement and Differential Features for High-Resolution Change Detection

Area of Science:

• Computer Science
• Artificial Intelligence
• Remote Sensing

Background:

• Semantic change detection is crucial for monitoring environmental and urban transformations.
• Existing methods struggle with class imbalance and variable object sizes, leading to misclassification and reduced accuracy.
• BGSNet provides a foundational framework but requires enhancements for improved performance.

Purpose of the Study:

• To propose SRDFNet (Semantic Refinement and Differential Features), a novel network for improved semantic change detection.
• To address limitations in existing methods, specifically class imbalance and variable object sizes.
• To enhance topological relationship perception and segmentation accuracy.

Main Methods:

• Introduced a hierarchical graph module (HGM) for compacting multi-scale features into semantic graph nodes, utilizing graph attention for semantic interaction.
• Developed a difference enhancement (DE) module with multi-receptive-field convolutions to extract difference information.
• Implemented a semantic refine (SR) module for lightweight residual refinement of bi-temporal semantic features.

Main Results:

• SRDFNet achieved state-of-the-art results on the SECOND and HRSCD datasets, outperforming BGSNet and other methods.
• On the SECOND dataset, SRDFNet achieved 87.64% OA and 70.31% mIoU, with notable gains over BGSNet.
• On the HRSCD dataset, SRDFNet achieved 98.13% OA and 52.67% mIoU, ranking first among evaluated methods.

Conclusions:

• SRDFNet effectively mitigates issues of class imbalance and variable object sizes in semantic change detection.
• The proposed HGM, DE, and SR modules contribute to enhanced topological perception and segmentation accuracy.
• SRDFNet represents a significant advancement in semantic change detection accuracy and robustness.