Multi-Scale Depthwise Separable Capsule Network for hyperspectral image classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification faces challenges in multi-scale feature extraction and pose information preservation.
• Existing methods often struggle with computational complexity and information loss.

Purpose of the Study:

• To propose a novel Multi-Scale Depthwise Separable Capsule Network (MDSC-Net) for improved HSI classification.
• To address the limitations of current HSI classification techniques regarding feature extraction and pose information.

Main Methods:

• MDSC-Net employs parallel multi-scale convolutional kernels for hierarchical feature extraction.
• Depthwise separable convolutions are utilized to reduce computational complexity.
• Independent capsule networks with dynamic routing process features of various scales to enhance translational invariance and preserve pose information.
• Features from different scales are concatenated for integrated multi-level analysis.

Main Results:

• MDSC-Net achieved high average accuracies: 94% on Kennedy Space Center, 98% on University of Pavia, and 99% on Salinas datasets.
• Demonstrated significant performance advantages over recent HSI classification models.
• Validated the effectiveness of the proposed MDSC-Net architecture.

Conclusions:

• MDSC-Net effectively extracts multi-scale features and preserves pose information for HSI classification.
• The proposed model offers an efficient and accurate solution for hyperspectral image analysis.
• MDSC-Net represents a significant advancement in the field of HSI classification.