Binary Change Guided Hyperspectral Multiclass Change Detection.

Area of Science:

• Remote Sensing
• Image Analysis
• Geospatial Information Science

Background:

• Hyperspectral imagery offers rich spectral information for detecting subtle changes.
• Existing hyperspectral binary change detection lacks fine class discrimination.
• Current hyperspectral multiclass change detection methods often neglect temporal correlation and accumulate errors.

Purpose of the Study:

• To propose an unsupervised network, BCG-Net, for enhanced hyperspectral multiclass change detection (HMCD).
• To improve both multiclass change detection and spectral unmixing performance by leveraging mature binary change detection techniques.
• To address limitations of existing HMCD methods, including error accumulation and lack of temporal correlation consideration.

Main Methods:

• Developed a novel partial-siamese united-unmixing module for multi-temporal spectral unmixing.
• Introduced a temporal correlation constraint guided by binary change detection pseudo-labels to refine the unmixing process.
• Implemented an innovative binary change detection rule robust to numerical variations and proposed iterative optimization for unmixing and change detection processes.

Main Results:

• BCG-Net achieved competitive or superior performance in multiclass change detection compared to state-of-the-art methods.
• The proposed network also demonstrated improved spectral unmixing results.
• Iterative optimization effectively reduced accumulated errors and bias between unmixing and change detection.

Conclusions:

• BCG-Net effectively enhances hyperspectral multiclass change detection by integrating binary change detection and temporal correlation.
• The method offers a significant advancement in simultaneously improving change detection and spectral unmixing.
• The proposed approach provides a robust solution for accurate HMCD in complex scenarios.