L₁ Sparsity-Regularized Attention Multiple-Instance Network for Hyperspectral Target Detection.

Area of Science:

• Remote Sensing
• Machine Learning
• Signal Processing

Background:

• Hyperspectral target detection often requires precise pixel-level labels, which are labor-intensive.
• Multiple-instance learning (MIL) offers a solution by using imprecise labels for entire areas (bags).
• Attention-based deep MIL models show promise but require further refinement for accurate target identification.

Purpose of the Study:

• To develop an L1 sparsity-regularized attention multiple-instance neural network (L1-attention MINN).
• To enhance hyperspectral target detection accuracy using imprecise training labels.
• To improve the discrimination of false positives in labeled data bags.

Main Methods:

• Proposed an L1-attention MINN model for hyperspectral target detection.
• Incorporated L1 sparsity regularization on attention weights for positive bags.
• Ensured compliance with MIL principles for improved discriminative ability.

Main Results:

• Achieved advanced performance on simulated and real-field hyperspectral datasets.
• Demonstrated superior results compared to state-of-the-art methods.
• Validated the effectiveness of the L1-attention MINN for imprecisely labeled data.

Conclusions:

• The L1-attention MINN is effective for hyperspectral target detection with imprecise labels.
• Sparsity constraints enhance the discriminative power of attention-based MIL models.
• The proposed method offers a robust solution for real-world hyperspectral imaging applications.