Deep clustering using 3D attention convolutional autoencoder for hyperspectral image analysis.

Area of Science:

• Remote Sensing
• Computer Vision
• Data Science

Background:

• Deep clustering is widely used in image and language processing.
• Hyperspectral image (HSI) processing faces challenges due to high dimensionality and complex spatial-spectral characteristics.
• Existing deep clustering methods may not be optimal for HSI data.

Purpose of the Study:

• To develop a specialized deep clustering model for hyperspectral image analysis.
• To address the challenges of high dimensionality and complex spatial-spectral features in HSIs.
• To improve the accuracy and efficiency of HSI clustering.

Main Methods:

• Dimensionality reduction using principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE).
• Feature extraction via a three-dimensional attention convolutional autoencoder (3D-ACAE) with a spatial-spectral attention mechanism.
• Clustering of compact data representations using an embedding and clustering layer.

Main Results:

• The proposed deep clustering model effectively reduces HSI dimensionality.
• The 3D-ACAE with attention mechanism successfully extracts enhanced spatial-spectral features.
• The model achieved superior performance in clustering HSIs across three public datasets.

Conclusions:

• The developed deep clustering approach is highly effective for HSI analysis.
• The integration of PCA, t-SNE, and 3D-ACAE offers a robust solution for HSI clustering.
• The model's superiority is validated by experimental results on diverse HSI datasets.