Kwantificatie van Heterogene Granulaire Mengsels met behulp van Hyperspectrale Deep Learning, met Toepassing op Maïskuil en Bulkhaver

With increasing global demands on food quality, safety, and production efficiency, the need for reliable, real-time quality control methods in the agrifood sector has grown significantly. Current practices for evaluating bulk granular products are often based on manual sampling and visual inspection, which are time-consuming, subjective, and provide limited representativeness of the entire batch. Moreover, these methods are typically performed offline, after production is complete, preventing timely feedback that could optimize processing conditions. Additionally, while distinguishing highly distinct objects is relatively easy, both for the human eye and computer vision systems, differentiating objects that are very similar in shape, size, and color (such as various grain types, tea blends, or granola) remains a significant challenge. To address these limitations, this study explores the integration of spectral imaging with deep learning-based computer vision for the automated quantification of heterogeneous granular mixtures. By using spectral imaging, which includes both multispectral and hyperspectral modalities, richer information than conventional RGB imaging can be leveraged by capturing data across a broader range of wavelengths. This allows for the differentiation of materials that appear visually similar but have distinct spectral signatures, enabling more precise classification and quantification of mixture components. In this research, we focused on two applications within the agrifood domain, both with high industrial relevance. The first case focuses on maize forage, a highly heterogeneous mixture consisting of leaves, stalks, husks, and kernels. Since only the kernels are relevant for assessing processing quality, and their degree of fragmentation directly impacts the digestibility of the feed or the efficiency of biogas production, this study evaluates the added value of spectral imaging for accurate kernel classification and quantification. The second use case addresses the contamination of bulk oats with spelt kernels, which are visually and spectrally very similar, making them difficult to distinguish using traditional methods. This is of particular importance as oats are naturally gluten-free, making accurate detection of gluten containing contaminants critical for meeting food safety standards. Here, a novel strategy was developed to overcome the bottleneck of ground truth annotation in hyperspectral datasets. This approach allowed for the efficient generation of training data without compromising spectral integrity, thereby enabling the development of robust deep learning models for quantification of objects in heterogeneous, granular mixtures.