Deep Ring-Block-Wise Network for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep learning excels in hyperspectral image (HSI) classification.
• Existing methods often neglect feature distribution, leading to suboptimal separability.
• Effective feature distribution requires block (intra-class compactness, inter-class separability) and ring (ring topology) properties.

Purpose of the Study:

• To propose a novel Deep Ring-Block-wise Network (DRN) for HSI classification.
• To enhance feature distribution by incorporating spatial geometric properties.
• To improve classification performance through more separable and discriminative features.

Main Methods:

• Developed a Deep Ring-Block-wise Network (DRN) integrating spatial geometry.
• Introduced a Ring-Block Perception (RBP) layer combining self-representation and ring loss.
• Designed an alternating update optimization strategy for the RBP layer.

Main Results:

• The proposed DRN method demonstrated superior classification performance.
• Features exported by DRN exhibited improved separability and discriminative power.
• Evaluated on Salinas, Pavia Centre, Indian Pines, and Houston datasets, DRN outperformed state-of-the-art approaches.

Conclusions:

• The DRN effectively considers feature distribution for HSI classification.
• The RBP layer successfully imposes block and ring properties on features.
• DRN offers a promising advancement in hyperspectral image classification accuracy.