Hyperspectral Image Classification Model Using Squeeze and Excitation Network with Deep Learning.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification is challenging due to complex data features, hindering traditional machine learning methods.
• Deep learning offers automatic feature learning, surpassing conventional hand-crafted approaches but faces difficulties with HSI's intricate data structure.
• Existing deep learning models require effective feature extraction strategies for optimal HSI classification performance.

Purpose of the Study:

• To propose an advanced deep feature extraction model specifically designed for hyperspectral image classification.
• To simultaneously extract spatial and spectral features from HSI data to improve classification system performance.
• To enhance the feature representation quality of Convolutional Neural Networks (CNNs) for HSI analysis.

Main Methods:

• Developed a novel Squeeze and Excitation Convolutional Neural Network (SE-CNN) model.
• Integrated the Squeeze and Excitation (SE) block with CNN architecture to refine feature extraction capabilities.
• Evaluated the SE-CNN model on three benchmark HSI datasets: Pavia Centre, Pavia University, and Salinas.

Main Results:

• The proposed SE-CNN model demonstrated superior performance in feature extraction and HSI classification.
• Achieved high overall accuracy: 96.05% (Pavia University), 98.94% (Pavia Centre), and 96.33% (Salinas).
• Outperformed established deep transfer learning models including VGG-16, Inception-v3, and ResNet-50 in both per-class and overall accuracy.

Conclusions:

• The SE-CNN model effectively extracts spatial and spectral features, significantly improving HSI classification accuracy.
• The integration of SE blocks enhances CNNs' ability to represent complex HSI data.
• This approach represents a significant advancement in deep learning-based hyperspectral image classification.