GhoMR: Multi-Receptive Lightweight Residual Modules for Hyperspectral Classification.

Area of Science:

• Computer Vision
• Remote Sensing
• Deep Learning

Background:

• Hyperspectral images (HSIs) possess numerous spectral bands, offering rich information for analysis.
• Convolutional Neural Networks (CNNs) are increasingly used for HSI analysis, leveraging learnable receptive fields (RFs) for feature extraction.

Purpose of the Study:

• To propose a novel multi-receptive CNN module (GhoMR) for efficient HSI classification.
• To address the issue of network weight increase with more RFs by integrating Ghost modules.

Main Methods:

• Developed the GhoMR module, incorporating multiple RFs within residual blocks for hierarchical feature extraction.
• Utilized Ghost modules to minimize feature redundancy and reduce network parameters.
• Constructed GhoMR-Net, a lightweight network employing GhoMR modules, and evaluated it on three public HSI datasets.

Main Results:

• GhoMR-Net achieved comparable or superior classification performance compared to ten state-of-the-art architectures.
• Performance was evaluated using Overall Accuracy (OA), Kappa coefficient (Kappa), and Average Accuracy (AA).
• The proposed lightweight architecture demonstrates effectiveness in HSI classification.

Conclusions:

• The GhoMR module offers an effective and efficient approach for HSI classification.
• GhoMR-Net provides a promising lightweight solution for analyzing complex hyperspectral data.
• The study's code is publicly available for further research and application.