Composite Neighbor-Aware Convolutional Metric Networks for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) struggle with hyperspectral image (HSI) classification due to limited local feature extraction.
• Graph Convolutional Networks (GCNs) can capture long-range dependencies but require computationally intensive full-batch training.
• Existing methods face challenges in effectively utilizing both local and non-local spectral-spatial information with limited labeled samples.

Purpose of the Study:

• To develop an efficient and effective method for supervised hyperspectral image classification.
• To overcome the limitations of CNNs and GCNs in spectral-spatial feature learning.
• To propose a novel network that leverages composite neighbors in a batchwise manner.

Main Methods:

• A composite neighbor-aware convolutional metric network (CNCMN) is proposed, integrating Euclidean and non-Euclidean neighbors.
• A composite convolution (CoConv) couples image and graph convolutions for adaptive feature extraction.
• A mini-batch metric classifier is introduced to optimize inter- and intra-class distances dynamically.

Main Results:

• The proposed CNCMN method demonstrates superior performance in hyperspectral image classification compared to existing mini-batch deep learning algorithms.
• Extensive experiments on three real-world hyperspectral datasets validate the effectiveness of the approach.
• The method achieves state-of-the-art performance, highlighting its advantages in spectral-spatial feature learning.

Conclusions:

• The CNCMN effectively extracts spectral-spatial features by considering composite neighbors in a batchwise manner.
• The proposed method offers a computationally efficient and high-performing solution for hyperspectral image classification.
• This work advances the field of hyperspectral image analysis by providing a novel and effective deep learning framework.