BSDR: A Data-Efficient Deep Learning-Based Hyperspectral Band Selection Algorithm Using Discrete Relaxation.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral band selection is vital for reducing high-dimensional data complexity.
• Attention-based algorithms are effective but require substantial training data due to numerous parameters.
• Existing methods face challenges with data efficiency and computational cost.

Purpose of the Study:

• To introduce Band Selection through Discrete Relaxation (BSDR), a data-efficient deep learning algorithm for hyperspectral band selection.
• To address the limitations of existing attention-based methods regarding parameter count and training data requirements.
• To enhance computational efficiency and analytical performance in hyperspectral data processing.

Main Methods:

• Developed Band Selection through Discrete Relaxation (BSDR), a novel deep learning approach.
• Implemented discrete relaxation to convert the band selection problem into a continuous optimization task.
• Focused on selecting target bands to minimize learnable parameters and data requirements.

Main Results:

• BSDR demonstrated superior performance in both regression and classification tasks on benchmark datasets.
• Achieved up to 25% and 34.6% accuracy improvements over attention-based and traditional algorithms, respectively.
• Reduced execution time by over 96.8%, indicating significant computational efficiency.

Conclusions:

• BSDR offers a highly effective and data-efficient solution for hyperspectral band selection.
• The algorithm overcomes the limitations of parameter-heavy methods, reducing training data needs and time.
• BSDR significantly enhances accuracy and efficiency in hyperspectral data analysis.