Asymmetric Weighted Logistic Metric Learning for Hyperspectral Target Detection

Area of Science:

• Remote Sensing
• Signal Processing
• Machine Learning

Background:

• Traditional target detection methods often assume Gaussian background distributions, limiting performance.
• Existing techniques struggle with imbalanced datasets, where target samples are scarce compared to background samples.

Purpose of the Study:

• To develop a novel target detection method overcoming limitations of traditional approaches.
• To address the challenges of non-Gaussian background distributions and imbalanced sample sizes in hyperspectral imagery.

Main Methods:

• A logistic metric-learning approach with a positive semidefinite constraint was formulated.
• An asymmetric weighted strategy was employed to manage imbalanced target and background samples.
• An accelerated proximal gradient method was utilized for optimization.

Main Results:

• The proposed asymmetric weighted logistic metric learning (AWLML) algorithm demonstrated superior performance.
• Experiments on three challenging hyperspectral datasets confirmed the effectiveness of AWLML.
• The method significantly improved upon state-of-the-art target detection results.

Conclusions:

• AWLML offers a robust solution for target detection in hyperspectral data.
• The method effectively handles complex background spectra and imbalanced sample distributions.
• AWLML represents a significant advancement in hyperspectral target detection capabilities.