A Supervised Segmentation Network for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep learning, particularly Convolutional Neural Networks (CNNs), is widely used for hyperspectral image (HSI) classification.
• Current CNN-based methods often crop HSI data into cubes, leading to complex spatial land-cover distributions and limited diversity, hindering generalization.

Purpose of the Study:

• To address the weak generalization capabilities of existing CNN-based HSI classification methods.
• To propose an end-to-end fully convolutional segmentation network (FCSN) for improved HSI classification.

Main Methods:

• Conducted experiments to validate the weak generalization of current CNN-based methods.
• Introduced a fine labeling strategy for detailed spatial land-cover distributions within HSI cubes.
• Developed a HSI cube generation method to enhance the diversity of spatial land-cover distributions.
• Proposed a FCSN to leverage spectral-spatial features from finely labeled HSI cubes.

Main Results:

• Demonstrated that recent CNN-based methods exhibit poor generalization.
• The proposed FCSN shows superior generalization capabilities when faced with changes in spatial land-cover distributions.
• Fine labeling and diverse HSI cube generation contribute to improved model performance.

Conclusions:

• The proposed FCSN effectively utilizes spectral-spatial features for HSI classification.
• Enhanced data diversity through fine labeling and generation improves model robustness.
• FCSN offers a promising approach for accurate and generalizable HSI classification.