Greedy Ensemble Hyperspectral Anomaly Detection.

Area of Science:

• Computer Vision
• Remote Sensing
• Data Science

Background:

• Hyperspectral images offer rich spectral information valuable for computer vision tasks.
• Anomaly detection in hyperspectral images is critical for identifying changes and abnormalities.
• Existing hyperspectral anomaly detection (HS-AD) algorithms have limitations due to varied background modeling assumptions.

Purpose of the Study:

• To develop an automated approach for selecting optimal HS-AD algorithms.
• To address the limitations of individual HS-AD algorithms in diverse scenarios.
• To improve the accuracy and reliability of anomaly detection in hyperspectral data.

Main Methods:

• Developed Greedy Ensemble Anomaly Detection (GE-AD), a two-stage stacking ensemble approach.
• Utilized a greedy search algorithm to select suitable base models from HS-AD and hyperspectral unmixing algorithms.
• Employed a supervised classifier in the second stage of the ensemble for final anomaly detection.

Main Results:

• GE-AD achieved statistically significant higher average F1-macro scores compared to individual and state-of-the-art ensemble methods.
• Demonstrated superior performance on multiple benchmark datasets, including ABU, San Diego, Salinas, Hydice Urban, and Arizona.
• GE-AD showed substantial improvements, outperforming previous methods by up to 28.53% on airport scenes.

Conclusions:

• The combination of greedy search and stacking ensemble offers an effective strategy for automated HS-AD model selection.
• GE-AD enhances anomaly detection accuracy and provides a robust solution for researchers with limited algorithm-specific knowledge.
• This work contributes to advancing hyperspectral anomaly detection and its practical applications.