Spectral Unmixing Benchmarks for Food Contamination Analytics

Abstract

Benchmark datasets play a critical role in advancing food contamination analytics because they provide common evaluation conditions for comparing detection models, spectral preprocessing pipelines, and unmixing algorithms. Meat contamination detection is particularly challenging when contaminant signals are weak, spatially limited, or spectrally similar to natural biological variation. Hyperspectral unmixing benchmarks enable researchers to evaluate whether models can separate weak contamination signatures from mixed tissue spectra. This topic focuses on the design and use of benchmark-driven spectral analytics for food safety inspection. Key issues include endmember selection, abundance estimation, spectral variability, labeling uncertainty, and reproducibility across acquisition settings. By emphasizing weak-signal detection, this literature direction supports food inspection systems that move beyond simple classification toward interpretable spectral decomposition and measurable contamination evidence.