Discriminative Multiple Instance Hyperspectral Target Characterization.

Area of Science:

• Remote Sensing
• Machine Learning
• Hyperspectral Imaging

Background:

• Accurate pixel-level labels for training data are often unavailable in hyperspectral imagery analysis.
• Sub-pixel targets are smaller than a single pixel, leading to mixed training data containing both target and non-target signatures.

Purpose of the Study:

• To present two novel methods, MI-SMF and MI-ACE, for discriminative multiple instance target characterization.
• To address the challenge of target detection using imprecisely-labeled and mixed training data in hyperspectral imagery.

Main Methods:

• Developed MI-SMF (Multiple Instance Spectral Matched Filter) and MI-ACE (Multiple Instance Adaptive Cosine Estimator) algorithms.
• These methods estimate a discriminative target signature from imprecisely-labeled and mixed training data.

Main Results:

• MI-SMF and MI-ACE demonstrate improved and consistent performance compared to existing multiple instance concept learning methods.
• The methods were evaluated on several hyperspectral sub-pixel target detection problems, showing enhanced detection capabilities.

Conclusions:

• The proposed MI-SMF and MI-ACE methods are effective for discriminative target characterization in hyperspectral imagery.
• These approaches offer a viable solution for sub-pixel target detection when dealing with limited or imprecise training data.