Robust Spatial-Spectral Squeeze-Excitation AdaBound Dense Network (SE-AB-Densenet) for Hyperspectral Image...

Area of Science:

• Artificial Intelligence
• Remote Sensing
• Computer Vision

Background:

• Deep learning has advanced hyperspectral image (HSI) classification.
• Common convolutional neural networks (CNNs) face challenges with HSI data complexity and computational cost due to numerous parameters.
• Extracting diverse spatial-spectral features further complicates CNN architectures.

Purpose of the Study:

• To design an optimized deep learning model for HSI classification that emphasizes significant spatial-spectral features.
• To reduce computational complexity and improve classification performance in HSI data analysis.
• To enhance model stability and accuracy while mitigating overfitting.

Main Methods:

• Developed an optimized squeeze-excitation AdaBound dense network (SE-AB-DenseNet) integrating dense networks with AdaBound and squeeze-excitation modules.
• Employed the AdaBound optimizer to improve model stability and classification accuracy.
• Utilized the cutout regularization technique to address overfitting in HSI spatial-spectral classification.

Main Results:

• The SE-AB-DenseNet model demonstrated competitive classification accuracy on the Indian Pines and Salinas hyperspectral datasets, even with limited training samples.
• The AdaBound optimizer contributed to an approximate 2% increase in classification accuracy.
• The model achieved high overall accuracies of 99.37% for Indian Pines and 99.78% for Salinas when combined with cutout regularization.

Conclusions:

• The proposed SE-AB-DenseNet with cutout regularization is an effective approach for HSI classification, offering improved accuracy and reduced computational cost.
• The integration of AdaBound and squeeze-excitation modules enhances model performance and stability.
• This method provides a robust solution for analyzing hyperspectral data, outperforming state-of-the-art methods in specific scenarios.