Variational Multiple-Instance Learning With Embedding Correlation Modeling for Hyperspectral Target Detection.

Area of Science:

• Geoscience and Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral target detection relies heavily on spectral information but requires costly, high-quality target signatures or pixel-level data.
• Existing methods face challenges due to the difficulty and expense of obtaining precise supervised signals for target detection.

Purpose of the Study:

• To develop a weakly supervised hyperspectral target detection method that only requires region-level labels.
• To relax the dependency on rigid target priors like signatures or pixel-level annotations.
• To improve the accuracy and efficiency of target detection in hyperspectral imagery.

Main Methods:

• A variational multiple-instance neural network with embedding correlation modeling (VMIL-ECM) is proposed.
• The model uses region-level labels and models target locations as latent variables under a non-i.i.d. assumption.
• An expectation-maximization (EM) algorithm optimizes latent variables and learns spectral features, incorporating a transformer for instance embedding correlation and dynamic thresholding for supervised signals.

Main Results:

• VMIL-ECM demonstrates effectiveness across simulated and real-field hyperspectral datasets.
• The proposed method achieves state-of-the-art performance compared to existing techniques.
• The approach successfully estimates underlying ground-truth target locations using only region-level supervision.

Conclusions:

• VMIL-ECM offers a robust and effective solution for weakly supervised hyperspectral target detection.
• The method alleviates the need for extensive, costly data annotation.
• The publicly available code facilitates further research and application in remote sensing.