Convolutional Neural Network Based on Bandwise-Independent Convolution and Hard Thresholding for Hyperspectral Band...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification benefits from dimensionality reduction via band selection.
• Deep learning approaches for HSI band selection are emerging but often isolate selection from classification.
• Existing methods face challenges like loss of discriminative bands and high computational expense due to isolated or iterative training.

Purpose of the Study:

• To propose a novel Convolutional Neural Network (CNN) named BHCNN for end-to-end hyperspectral image classification.
• To integrate band selection, feature extraction, and classification into a single trainable network.
• To improve the efficiency and accuracy of HSI classification by optimizing spectral band selection.

Main Methods:

• Developed a BHCNN model incorporating bandwise-independent convolution and hard thresholding for band selection.
• Utilized bandwise 1x1 convolutions to process spectral bands independently, followed by hard thresholding to zero out unselected bands.
• Employed a straight-through estimator (STE) to optimize weights for unselected bands and a coarse-to-fine loss for interpretability.
• Integrated multiscale 3-D dilated convolutions for joint spatial-spectral feature extraction from selected bands.

Main Results:

• The BHCNN model effectively combines band selection, feature extraction, and classification in an end-to-end trainable network.
• Hard thresholding and STE successfully constrained and optimized spectral band selection.
• The proposed coarse-to-fine loss enhanced the interpretability of the STE.
• Experimental results on multiple HSI datasets showed superior classification performance compared to state-of-the-art methods using selected spectral bands.

Conclusions:

• The proposed BHCNN method offers an effective and efficient solution for hyperspectral image classification.
• Integrating band selection within the deep learning framework leads to improved classification accuracy.
• BHCNN demonstrates the potential of end-to-end trainable networks for advanced HSI analysis.