Going Deeper With Contextual CNN for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image classification is crucial for analyzing detailed spectral information.
• Existing deep convolutional neural networks (CNNs) have limitations in exploiting local contextual information.
• Integrating spatial and spectral features effectively remains a challenge in hyperspectral image analysis.

Purpose of the Study:

• To introduce a novel, deeper, and wider deep convolutional neural network (CNN) for hyperspectral image classification.
• To develop a method that optimally explores local contextual interactions by jointly exploiting spatio-spectral relationships.
• To improve the accuracy and efficiency of hyperspectral image classification.

Main Methods:

• A contextual deep CNN architecture is proposed, featuring a multi-scale convolutional filter bank.
• The filter bank jointly exploits local spatial and spectral relationships of neighboring pixel vectors.
• A fully convolutional network processes the combined spatio-spectral feature map for pixel-wise classification.

Main Results:

• The proposed contextual deep CNN demonstrates enhanced classification performance.
• Superior results were achieved compared to current state-of-the-art methods.
• The approach was validated on three benchmark hyperspectral datasets: Indian Pines, Salinas, and University of Pavia.

Conclusions:

• The novel contextual deep CNN effectively integrates local spatio-spectral information for improved hyperspectral image classification.
• The proposed method offers a significant advancement over existing CNN-based approaches.
• The enhanced performance on benchmark datasets validates the efficacy of the joint spatio-spectral feature exploitation.