Hyperspectral image analysis : unveiling new perspectives in representation learning techniques

Hyperspectral imaging (HSI) combines digital imaging with spectroscopy, capturing hundreds of contiguous wavelength bands per pixel. This produces a three-dimensional “hypercube” that integrates spatial and spectral information, offering powerful opportunities for material analysis, environmental monitoring, and beyond. Yet, HSI also poses challenges: high dimensionality, limited annotated data, and heavy computational demands. My thesis addresses these challenges through three complementary approaches. First, I develop a deep learning framework for multi-label classification that demonstrates strong performance in complex, mixed-material scenes and highlights limitations in common annotation practices. Second, I explore self-supervised contrastive learning to enhance classification accuracy under limited supervision, demonstrating substantial gains across diverse datasets. Third, I apply explainability-driven dimensionality reduction to identify the most informative spectral bands, reducing redundancy while maintaining or even improving accuracy. Together, these contributions demonstrate that tailored representation learning strategies and explainability-driven dimensionality reduction can deliver hyperspectral classifiers that are accurate, computationally efficient, and adaptable to challenging data conditions. The results highlight practical pathways for improving hyperspectral image analysis and open opportunities for further exploration across diverse application domains.