Convolutional neural network for hyperspectral data analysis and effective wavelengths selection.

Area of Science:

• Hyperspectral Imaging
• Machine Learning
• Data Analysis

Background:

• Spectral-spatial information fusion improves near-infrared hyperspectral data analysis.
• Existing methods often require complex pipelines and expert parameter tuning.
• Convolutional Neural Networks (CNNs) offer automated feature representation.

Purpose of the Study:

• To develop an automated spectral-spatial classification method for hyperspectral data.
• To introduce a two-branch CNN (2B-CNN) for improved classification and effective wavelengths (EWs) selection.
• To evaluate the 2B-CNN's performance against traditional methods.

Main Methods:

• A two-branch convolutional neural network (2B-CNN) architecture was designed.
• The network was trained and tested on diverse datasets: herbal medicine, coffee bean, and strawberry.
• Learned weights from the 2D branch were used for effective wavelengths (EWs) selection.

Main Results:

• The 2B-CNN achieved an average classification accuracy of 96.72%, outperforming SVM, 1D CNN, and GLCM-SVM.
• EWs selected using the 2B-CNN's weights yielded an average accuracy of 96.02%, surpassing the Successive Projections Algorithm.
• The proposed method effectively integrates spectral and spatial discriminative power.

Conclusions:

• The 2B-CNN provides a general, automated approach for spectral-spatial classification of hyperspectral data.
• The method offers an efficient and effective way to select informative wavelengths.
• This approach reduces reliance on manual parameter optimization and expert knowledge.