Transform-based signal representations for efficient AI-based hyperspectral imaging

This thesis concentrates on the formulation of transform-based signal representations to enable efficient AI-based classification of hyperspectral images. The doctoral program constitutes an industrial initiative originating from the HYPER ABC project (HYPERspectral imaging through Artificial intelligence for Building Control) funded by Regione Lazio and Superelectric srl within the doctoral funding program PO FSE 2014-2020, aimed at developing advanced and innovative artificial intelligence techniques for integration into monitoring systems. The thesis addresses the escalating complexity of employing hyperspectral images within industrial applications, emphasizing the critical necessity for novel methodologies to enhance operational efficiency in terms of processing speed and classification accuracy. Hyperspectral imaging measures the spatial and spectral characteristics of an object at different wavelengths and provides big data-rich information; on the other side, hyperspectral images are high dimensional data that require high computational resources to be processed. The thesis explores innovative methodologies aimed at enhancing hyperspectral image classification through optimization techniques applied to both input data and network architecture, in line with the industrial aim of operational efficiency. The investigation starts by optimizing the input data through the application of dimensionality reduction methods for extracting relevant information, removing redundancies, and compressing data. This optimization involves feature extraction and selection techniques combined with entropy-based and wavelet-based approaches. Regarding feature extraction, Principal Components Analysis and Discrete Wavelet Transform are employed to transform data and extract crucial information through a linear approach, while entropy-based methodologies automatically define the type and number of features to retain. Alternatively, Continuous Wavelet Transform and Wavelet Leaders are employed for feature selection through a non-uniform sampling to automatically select informative bands without compromising their intrinsic physical meaning. Experimental results demonstrate the effectiveness of these methodologies in efficiently reducing hyperspectral data while maintaining or even improving classification accuracy compared to that achieved when using all original bands. AI-based classifiers, specifically Support Vector Machine and Convolutional Neural Network, are selected for evaluating the proposed methods. To further exploit signal representation, a preliminary study for optimizing neural network architecture for hyperspectral image classification is also presented. A Heisenberg-based method is proposed for identifying a rule for the size of cascaded filters of the convolutional layers of a Convolutional Neural Network that leads to higher accuracy in a suitable time. Lastly, the research investigates the implementation of these techniques within industrial environments. This involves empirical results from real-world data analysis in several fields, including assessments on both multispectral and hyperspectral images. These findings show the applicability and effectiveness of the proposed methodologies within industrial domains, presenting a comprehensive approach to enhance hyperspectral image classification with advanced dimensionality reduction and artificial intelligence techniques while optimizing operational workflows.