Sustainability-focused AI solutions and data-driven methods for dimensionality reduction of hyperspectral images

Hyperspectral imaging (HSI) is a non-invasive and non-contact technology that identifies the composition of an object based on the analysis of hundreds of wavelengths reflected from it, visualizing their spectral and spatial distributions. This high spectral resolution allows hyperspectral images (HI) to reveal detailed information about objects that are invisible to the human eye, making it a powerful tool for industrial applications. However, the high dimensionality of HI data demands significant computational resources, especially when integrated into Machine Learning (ML) and Deep Learning (DL) frameworks for sustainability-focused artificial intelligence (AI) solutions. Advanced dimensionality reduction techniques play a crucial role in the deployment of efficient and sustainable AI-based systems in industrial settings. To address the dimensionality reduction challenge in HSI, two primary approaches are proposed in this work. The first approach requires non-linear approximation methods within an appropriate expansion basis for selecting significant spectral features in a transformed domain, while the second approach benefits from non-uniform sampling methods for selecting relevant spectral bands from the original spectrum. The efficacy and benefits of these methods are examined through comparative studies using benchmarking datasets, demonstrating their potential to provide high classification accuracy while significantly reducing computational load and processing time.