Hyperspectral remote sensing image retrieval system using spectral and texture features.

Area of Science:

• Remote Sensing
• Image Analysis
• Computer Vision

Background:

• Content-based image retrieval (CBIR) systems are prevalent, yet few address hyperspectral remote sensing images.
• Hyperspectral images present unique challenges, including mixed pixels, requiring specialized feature extraction and similarity measures.

Purpose of the Study:

• To propose a novel hyperspectral remote sensing image retrieval system.
• To enhance retrieval performance by integrating spectral and texture features.
• To address limitations in existing hyperspectral image analysis schemes.

Main Methods:

• Extracted endmembers as spectral features using an improved automatic pixel purity index algorithm to account for mixed pixels.
• Extracted texture features using the gray level co-occurrence matrix.
• Designed a mixed similarity measurement combining spectral information divergence and spectral angle match for spectral features, and Euclidean distance for textural features.
• Synthesized true color images for result visualization and optimized retrieval via user relevance feedback adjusting feature weights.

Main Results:

• The proposed system demonstrated superior retrieval performance compared to existing hyperspectral analysis schemes on NASA datasets.
• The integration of spectral and texture features effectively improved retrieval accuracy.
• User relevance feedback allowed for adaptive optimization of retrieval results.

Conclusions:

• The developed hyperspectral image retrieval system offers a robust and effective solution for analyzing hyperspectral remote sensing data.
• The system's ability to handle mixed pixels and incorporate user feedback provides a significant advancement in the field.
• The proposed methods show promise for practical applications in remote sensing image retrieval.