Hyperspectral Image Classification with Capsule Network Using Limited Training Samples.

Area of Science:

• Computer Science
• Artificial Intelligence
• Remote Sensing

Background:

• Deep learning, particularly Convolutional Neural Networks (CNNs), has significantly advanced hyperspectral image (HSI) classification.
• Capsule Networks (CapsNets) have emerged as a novel architecture to enhance CNN performance.
• Limited training samples pose a challenge for robust HSI classification models.

Purpose of the Study:

• To introduce a modified two-layer CapsNet for HSI classification using limited training data.
• To evaluate the robustness and representation capabilities of CapsNets compared to CNNs.
• To investigate the performance of CapsNets on both complex and simple HSI datasets.

Main Methods:

• A modified two-layer Capsule Network (CapsNet) architecture was designed for HSI classification.
• The CapsNet model was trained and evaluated on two real-world HSI datasets: PaviaU (complex) and SalinasA (simple).
• A comparative analysis was conducted against state-of-the-art CNNs, Random Forests (RFs), and Support Vector Machines (SVMs).

Main Results:

• CapsNet demonstrated superior accuracy and convergence behavior compared to CNNs on the complex PaviaU dataset.
• For the PaviaU dataset, CapsNet achieved a Kappa coefficient of 0.9456, overall accuracy of 95.90%, and average accuracy of 96.27%.
• CapsNet exhibited higher confidence in predicted probabilities, supported by probability map and uncertainty analyses.

Conclusions:

• The proposed CapsNet offers significant advantages for HSI classification, particularly under limited training sample conditions.
• CapsNets provide a promising alternative to CNNs, delivering improved performance and higher confidence in classification results.
• CapsNet outperforms traditional machine learning classifiers (RFs, SVMs) and state-of-the-art CNNs for HSI classification tasks.