Multiple Instance Learning for Multiple Diverse Hyperspectral Target Characterizations.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral target characterization faces challenges with imprecise labels, mixed data, and multiple target types in real-world scenarios.
• Lack of accurate pixel-level annotations and the presence of subpixel or occluded targets complicate traditional methods.

Purpose of the Study:

• To develop a novel hyperspectral target characterization method robust to imprecise training data.
• To generate diverse multiple hyperspectral target signatures using a multiple instance learning (MIL) framework.

Main Methods:

• The proposed method employs a multiple instance learning (MIL) framework, utilizing only bag-level training samples and labels.
• A multiple characterization MIL formulation with a diversity-promoting term is introduced to learn a set of distinct target signatures.

Main Results:

• The method successfully addresses issues of mixed data and lack of pixel-level labels by using bag-level information.
• It effectively learns diverse target signatures, overcoming challenges posed by multiple target types within training samples.

Conclusions:

• The developed hyperspectral target characterization method demonstrates effectiveness in producing diverse target signatures.
• Experimental results on synthetic and real-world data confirm the method's capability in hyperspectral target detection.