TBSSF-Net: three-branch spatial-spectral fusion network for hyperspectral image classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Data Science

Background:

• Hyperspectral images (HSI) offer rich spatial and spectral information, making them valuable for various applications.
• HSI classification is a key research area, often relying on spatial-spectral feature fusion.
• Traditional methods using large neighborhood windows can lead to training and test dataset independence issues.

Purpose of the Study:

• To propose a novel network, the three-branch spatial-spectral fusion network (TBSSF-Net), for HSI classification.
• To address the challenge of data independence in HSI classification by utilizing smaller patch sizes.
• To enhance classification performance by effectively fusing spatial and spectral information.

Main Methods:

• Developed a three-branch network architecture: spatial key details aggregation, spatial semantic knowledge refinement, and spectral band signal granularity.
• Employed smaller patch sizes to maintain independence between training and testing datasets.
• Integrated spatial and spectral feature extraction within a unified framework.

Main Results:

• The TBSSF-Net demonstrated superior and effective performance on four public HSI datasets.
• Achieved significant classification performance even with limited training data.
• The network successfully retained spatial details and captured global semantic context while refining spectral information.

Conclusions:

• The proposed TBSSF-Net offers an effective solution for hyperspectral image classification.
• The spatial-spectral fusion approach with smaller patch sizes enhances classification accuracy and robustness.
• TBSSF-Net shows promise for diverse HSI classification tasks with varying training set sizes.