A Robust Context-Based Deep Learning Approach for Highly Imbalanced Hyperspectral Classification

Area of Science:

• Hyperspectral imaging
• Deep learning
• Remote sensing

Background:

• Deep convolutional neural networks (DCNs) are state-of-the-art for hyperspectral image classification.
• However, DCNs underperform on noisy hyperspectral datasets.

Purpose of the Study:

• To propose a feature augmentation approach to enhance noise resistance in imbalanced hyperspectral classification.
• To evaluate the performance of a context-based DCN against standard DCN and PCA+DCN models.

Main Methods:

• A novel feature augmentation method calculating context-based features was developed.
• The proposed method utilizes a deep convolutional neural network (DCN).
• Experiments were conducted on Pavia datasets, comparing DCN, PCA+DCN, and the context-based DCN on both original and noisy data.

Main Results:

• Standard DCN and PCA+DCN performed well on clean data but poorly on noisy data.
• The proposed context-based DCN significantly outperformed other models in the presence of noise.
• Comparable classification accuracy was maintained on clean hyperspectral images.

Conclusions:

• The context-based feature augmentation approach significantly improves the robustness of DCNs for hyperspectral image classification in noisy conditions.
• This method offers a promising solution for reliable hyperspectral data analysis where noise is a concern.