Hybrid spatial-spectral generative adversarial network for hyperspectral image classification.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Hyperspectral image (HSI) classification relies on extracting both spectral and spatial features.
• 3D Convolutional Neural Networks (CNNs) excel at joint spatial-spectral feature extraction but are computationally intensive.
• Generative Adversarial Networks (GANs) show promise for HSI classification but often use 2D CNNs, limiting feature extraction.

Purpose of the Study:

• To propose a novel hybrid spatial-spectral generative adversarial network (HSSGAN) for effective HSI classification.
• To address the computational complexity of 3D CNNs while leveraging their feature extraction capabilities.
• To enhance classification accuracy, particularly in scenarios with limited training samples.

Main Methods:

• Developed a hybrid CNN structure for both generator and discriminator in the GAN.
• Utilized a 3D CNN in the discriminator for multi-band spatial-spectral feature extraction, followed by a 2D CNN for spatial representation.
• Incorporated a channel and spatial attention mechanism (CSAM) to enhance spectral features and suppress redundant spatial information.

Main Results:

• The proposed HSSGAN demonstrated satisfactory classification performance on four widely used hyperspectral datasets.
• The hybrid approach effectively balances feature extraction and computational complexity.
• Significant improvements were observed compared to conventional methods, especially when using few training samples.

Conclusions:

• The HSSGAN is an effective deep learning model for hyperspectral image classification.
• The integration of 3D CNNs and attention mechanisms enhances the model's ability to capture complex spatial-spectral features.
• The HSSGAN offers a viable solution for HSI classification challenges, particularly in data-scarce environments.