Year: 2026
Type: preprint
Abstract: Hyperspectral imaging (HSI) combines spatial and spectral information, producing high-dimensional datasets for which machine learning (ML) methods have become a key tool for classification. However, publicly available HSI datasets that explicitly require the joint exploitation of both spatial and spectral dimensions for classification remain limited. Here, we present HyperCookies, a carefully annotated HSI dataset comprising images of 160 cookies belonging to four classes defined by both spectral and spatial characteristics. Spectral variation is introduced through chocolate droplets with different cocoa contents, and spatial variation is represented by two topping arrangements. HSI data were acquired using two complementary line-scan cameras covering the visible and near-infrared (VNIR, 400–1000 nm) and short-wave infrared (SWIR, 900–1700 nm) ranges. The cookies were imaged both before and after baking, enabling the study of processing-induced changes. Baseline experiments using both spectral ML classifiers and image-based convolutional neural networks demonstrate that neither spectral nor spatial information alone is sufficient to reliably discriminate all four classes. The HyperCookies dataset is therefore made publicly available as a challenging and realistic benchmark for the development and evaluation of methods that jointly exploit spectral and spatial information in hyperspectral images.
Source: Research Square
Institution: KU Leuven
Language: English
Subfield: Analytical Chemistry
Field: Chemistry
Domain: Physical Sciences
Open Access status: green
Funder: KU Leuven
OpenAlex ID: W7225329823