Matrix Profile for anomaly detection in hyperspectral images: Application to the early detection of apple scab

Hyperspectral imaging has become a key analytical tool for the non-destructive characterization of complex materials, providing both spatial and spectral information. However, the detection of anomalies in hyperspectral images remains challenging due to high dimensionality, strong spectral correlations and the absence of prior knowledge on the nature of potential defects or outliers. Robust, data-driven methods that do not rely on supervised learning are therefore of great interest in analytical chemistry. In this work, we propose a novel transformation based on the Matrix Profile to enhance local spectral dissimilarities in hyperspectral images, facilitating anomaly detection in a subsequent analysis step. Originally developed for time series analysis, the Matrix Profile provides an efficient and exact measure of similarity between subsequences, enabling the identification of unusual patterns without assuming a specific data distribution. By reformulating hyperspectral data as collections of spectral sequences, the proposed approach allows spatial and spectral anomalies to be detected in an unsupervised and computationally efficient manner. The method is evaluated on a hyperspectral dataset representative of analytical imaging contexts, and its ability to highlight chemically or physically abnormal regions is assessed. Results demonstrate that the Matrix Profile–based approach successfully identifies subtle anomalies that may be overlooked by conventional distance- or variance-based techniques. This work highlights the potential of Matrix Profile methods as a robust and scalable tool for exploratory analysis and quality assessment in hyperspectral analytical imaging. • A novel application of the Matrix Profile for anomaly detection in hyperspectral images. • Fully unsupervised detection of spectral and spatial anomalies. • No assumption on data distribution or anomaly type. • Efficient and scalable method suitable for high-dimensional hyperspectral data. • Promising tool for exploratory analysis and quality control in analytical imaging.