Related Experiment Video
Updated: Jun 29, 2025
07:05
Applying Hyperspectral Reflectance Imaging to Investigate the Palettes and the Techniques of Painters
Published on: June 18, 2021
2.4K
Advancing Hyperspectral Image Analysis with CTNet: An Approach with the Fusion of Spatial and Spectral Features.
Dhirendra Prasad Yadav1,2, Deepak Kumar2, Anand Singh Jalal1
1Department of Computer Engineering & Applications, G.L.A. University, Mathura 281406, Uttar Pradesh, India.
Sensors (Basel, Switzerland)
|March 28, 2024
Summary
This study introduces a novel Convolutional Transformer Network (CTNet) to improve hyperspectral image classification by enhancing spatial and spectral features. The method achieves superior performance, demonstrating its effectiveness over existing techniques.
Area of Science:
- Remote Sensing
- Computer Vision
- Machine Learning
Background:
- Hyperspectral image classification faces challenges due to high data dimensionality and limited spatial resolution.
- Existing methods struggle with insufficient data samples and poor spatial detail.
Purpose of the Study:
- To enhance spatial and spectral features for improved hyperspectral image classification.
- To address limitations of data samples and spatial resolution in hyperspectral datasets.
Main Methods:
- A two-scale module-based Convolutional Transformer Network (CTNet) is proposed.
- Module 1: Virtual RGB image creation using a pre-trained ResNeXt model for spatial feature enhancement.
- Module 2: Principal Component Analysis (PCA) for dimension reduction and Enhanced Attention-based Vision Transformer (EAVT) with a multiscale attention mechanism for spectral feature enhancement.
- A joint module fuses spatial and spectral features to generate an enhanced feature vector.
Main Results:
- The proposed CTNet demonstrates superior performance compared to state-of-the-art methods.
- Achieved high Average Accuracy (AA) values: 97.87% (PU), 97.46% (PUC), 98.25% (SV), and 84.46% (Houston13).
Conclusions:
- The CTNet effectively enhances both spatial and spectral features for hyperspectral image classification.
- The proposed method offers a significant advancement in handling high-dimensional and spatially limited hyperspectral data.
