Shared subspace learning via partial Tucker decomposition for hyperspectral image classification.

Area of Science:

• Hyperspectral Imaging
• Machine Learning
• Multidimensional Data Analysis

Background:

• Hyperspectral imaging generates complex, high-dimensional data.
• Traditional methods often flatten data, losing crucial multidimensional relationships.
• Image-level labels require methods that handle spatially heterogeneous phenomena.

Purpose of the Study:

• To introduce a novel tensor-based classification framework, Shared Subspace Tensor Classification (SSTC).
• To address challenges in hyperspectral image analysis, particularly for heterogeneous sample distributions.
• To enable effective dimensionality reduction and feature extraction for classification tasks.

Main Methods:

• Utilized partial Tucker decomposition to learn shared spatial and spectral subspaces.
• Employed core tensors for discriminative feature extraction from hyperspectral data.
• Applied the framework to food quality assessment tasks: plum bruising detection and mango ripeness classification.

Main Results:

• SSTC achieved competitive performance against deep learning methods in plum bruising detection, with superior interpretability and efficiency.
• The framework significantly outperformed existing techniques in mango ripeness classification, especially with limited training data.
• Learned decompositions revealed physically meaningful patterns, demonstrating interpretable feature extraction.

Conclusions:

• SSTC provides an effective and interpretable tensor-based approach for hyperspectral image classification.
• The framework offers efficient data compression while maintaining or improving classification accuracy.
• SSTC demonstrates significant advantages in food quality assessment applications, particularly with limited data scenarios.