Correlation awareness evolutionary sparse hybrid spectral band selection algorithm to detect aflatoxin B1 contaminated almonds using hyperspectral images

Abstract

Aflatoxin B1 is a harmful metabolite that frequently contaminates almonds, other nuts, and grains. Prolonged consumption of foods contaminated with aflatoxin B1 can lead to severe health issues. Hyperspectral imaging enables rapid, non-destructive detection of aflatoxin B1, but its high dimensionality complicates data analysis and increases complexity of classification models. This paper presents a novel hybrid spectral band selection algorithm designed to classify aflatoxin B1 in almonds, suitable for industrial applications. The algorithm operates in two main steps. Firstly, it identifies significant spectra individually based on various tree-based boosting ensemble techniques and multilayer perceptron networks. Then, the significant spectra were optimized using the correlation-aware sparse spectral band selection process. The proposed algorithm was evaluated on three hyperspectral image datasets and was compared with existing classical methods. The selected 4 to 10 spectra achieved comparable classification accuracy compared to the full spectra model and can be used in industrial applications.