Shared subspace learning via partial Tucker decomposition for hyperspectral image classification

A tensor-based classification framework, which we refer to as Shared Subspace Tensor Classification (SSTC), is proposed for hyperspectral imaging applications where image-level labels must predict phenomena that distribute heterogeneously across samples. Instead of flattening the natural multi-dimensional structure of hyperspectral data, our approach employs partial Tucker decomposition to learn shared spatial and spectral subspaces across samples, enabling effective dimensionality reduction while preserving crucial relationships between dimensions. Core tensors encoding each sample's projection onto these subspaces provide discriminative features that achieve strong classification performance even with simple classifiers. We evaluate the framework on two food quality assessment tasks: detecting subsurface bruising in plums and classifying mango ripeness. Our method demonstrates competitive performance compared to deep learning approaches while offering superior interpretability and computational efficiency for plum bruising detection. In mango ripeness classification, where limited training data poses challenges for deep learning, our approach substantially outperforms existing techniques. Analysis of the learned decomposition reveals physically meaningful patterns aligned with domain knowledge, demonstrating both effective classification and interpretable feature extraction. The framework provides efficient data compression while maintaining or improving classification accuracy compared to traditional approaches.