Concept Disentangled Transformer for Hyperspectral Image Classification
Year: 2026
Type: article
Abstract: Hyperspectral image (HSI) classification remains challenging due to high spectral dimensionality, extremely limited labeled samples, and complex spatial–spectral dependencies. Recent Transformer-based models provide strong contextual modeling, yet their high-dimensional latent representations are often weakly structured under low-label regimes, which can lead to unstable class separation and limited analyzability of the decision process. This article proposes CDFormer (CDF), a Transformer framework that introduces an explicit concept bottleneck (CD) to impose a structured intermediate representation for classification. CDF projects attention-enhanced token features into a low-dimensional concept space composed of structured latent coordinates for decision decomposition and predicts labels through a linear class-concept readout, yielding an exact logit decomposition into concept-wise contributions rather than predefined semantic attributes. To improve representation learning under scarce supervision, we train CDF using a composite objective that combines standard classification loss with orthogonality and sparsity regularization (SP) and a supervised contrastive term applied in concept space, encouraging reduced feature redundancy, selective coordinate activation, and improved intraclass compactness/interclass separation. Experiments on five benchmark HSI datasets demonstrate that CDF achieves consistent performance compared with recent state-of-the-art baselines.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7169767827