Machine learning methods for hyperspectral imaging: from reconstruction to classification

Abstract

With the growing demands for efficient and scalable food analysis, agriculture, and healthcare applications, the need for improved data acquisition and processing techniques has become increasingly significant. Near Infrared Spectroscopy has emerged as a powerful tool for non-invasive analysis in these domains, providing key insights into material composition at a spectral level. Hyperspectral imaging, in particular, captures a wide range of wavelengths, making it ideal for analysing complex materials and substances that standard RGB or Near-Infrared imaging may miss. However, traditional hyperspectral imaging systems are costly, computationally demanding, third party locked, and are often impractical for edge deployment. This thesis investigates novel techniques for augmenting Near-Infrared spectroscopy and hyperspectral imaging through advanced machine learning methods, specifically tailored for edge computing platforms. By leveraging machine learning models, this work enhances the resolution of spectral data, allowing for higher precision in tasks such as classification and reconstruction without relying on expensive hardware. The research further explores the application of RGB-to-hyperspectral reconstruction techniques, enabling the conversion of standard RGB images into hyperspectral ones, this democratizing access the hyperspectral analysis. Additionally, this thesis contributes to the field through several case studies, including the classification of cow diets based on Mid-Infrared spectra, hyperspectral image reconstruction for food science, and the development of innovative architectures for spectral and spatial segmentation and classification. The contributions, published in peer-reviewed conferences and journals, demonstrate the effectiveness of the proposed methods in diverse applications, from agricultural data analysis to food imaging. By integrating hyperspectral machine learning processing and edge computing, this research not only advances the field of Near-Infrared spectroscopy but also broadens the scope and accessibility of hyperspectral imaging. The findings of this work have broad implications for industries such as food processing, agriculture, and healthcare, where rapid and accuracy spectral analysis is critical for decision-making and quality control. Thesis is embargoed until 31 July 2027.