Learning Deep Hierarchical Spatial-Spectral Features for Hyperspectral Image Classification Based on Residual 3D-2D...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral imaging captures detailed spectral information in numerous narrow bands.
• Pixel-wise classification is crucial for hyperspectral applications, but deep learning faces challenges due to limited annotated data.
• The 'small-sample problem' hinders the effectiveness of Convolutional Neural Networks (CNNs) in hyperspectral classification.

Purpose of the Study:

• To develop an optimized deep learning model for hyperspectral image classification with limited training samples.
• To address the challenges posed by the scarcity of labeled data in hyperspectral datasets.
• To improve the accuracy and efficiency of hyperspectral image analysis.

Main Methods:

• Designed an 11-layer CNN model named R-HybridSN (Residual-HybridSN).
• Integrated 3D-2D-CNN, residual learning, and depth-separable convolutions for enhanced feature learning.
• Evaluated the model's performance on three public hyperspectral datasets (Indian Pines, Salinas, University of Pavia) using minimal training data.

Main Results:

• R-HybridSN achieved high classification accuracies: 96.46% on Indian Pines, 98.25% on Salinas, and 96.59% on University of Pavia.
• These results were obtained using only 5%, 1%, and 1% labeled data, respectively.
• The model significantly outperformed existing contrast models in limited-sample scenarios.

Conclusions:

• R-HybridSN effectively learns deep hierarchical spatial-spectral features from limited hyperspectral data.
• The proposed network optimization strategies overcome the 'small-sample problem' in hyperspectral classification.
• R-HybridSN offers a promising solution for accurate hyperspectral image analysis with minimal annotations.