CGMNet: A Center-Pixel and Gated Mechanism-Based Attention Network for Hyperspectral Change Detection

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Computer Vision

Background:

• Deep learning, especially CNNs and transformers, has advanced hyperspectral image change detection (HSI CD).
• Existing models may include irrelevant spatial information, hindering HSI CD accuracy.
• A robust method is needed to effectively utilize spectral-spatial features for HSI CD.

Purpose of the Study:

• To propose a novel attention network, CGMNet, for enhanced HSI CD.
• To leverage center-pixel significance and gated mechanisms for improved accuracy and robustness.
• To introduce a new benchmark dataset for coastal remote sensing research.

Main Methods:

• A center-pixel and gated mechanism-based attention network (CGMNet) was developed.
• Key components include gated-based center spatial attention (GCSA) and gated-based spectral attention (GSA) modules.
• A global transform fusion (GTF) module integrates features, and a new Hangzhou Bay (HZB) dataset was created.

Main Results:

• CGMNet demonstrated superior performance in HSI CD across multiple datasets, including the new HZB dataset.
• The proposed GCSA and GSA modules effectively emphasize relevant spatial and spectral features.
• The GTF module successfully fused multi-level features for comprehensive analysis.

Conclusions:

• CGMNet offers a significant advancement in HSI CD accuracy and robustness.
• The proposed attention mechanisms provide effective feature selection and fusion.
• The HZB dataset will facilitate future research in coastal remote sensing and HSI CD.