Explainable AI-assisted hybrid self-organising maps and deep learning algorithms for detecting pistachio adulteration with peas

Abstract

In this study samples of raw pistachio nuts were prepared containing from 0% to 60% adulteration with dried peas. This training set of samples were analysed by Raman spectroscopy at frequencies from 200 to 2000 cm-1 and an integrated AI model was used to detect this adulteration using 18,000 laboratory data points. Novel hybrid Self-organising Maps (SOM) and Convolutional Neural Network (CNN) methodologies were developed to intuitively examine pistachio adulteration with peas. SOM, an unsupervised learning (UL) method, was used to convert the data into lower-dimensional digital images. The numerical values depicted in the images showed the extent of pure and contaminated pistachio nuts in the mixtures. Image classification is a key subject in computer vision which aims to distinguish multiple clusters based on distinct features of images. Since image pixel values arranged in a two-dimensional grid and hexagonal lattices show similarities among pistachio adulteration characteristics, CNN was employed to efficiently sort and classify images for detecting pistachio-pea features. It involves determining a function that correlates the pixel intensities of an image with a specified class on the lattices. The outcomes of training and validation were satisfactory. This innovative hybridisation method is distinguished by its ability to cluster and represent high-dimensional unsupervised datasets as two-dimensional images, while also distinguishing between fraudulent and authentic pistachios. The CNN was constructed with a 4-layer architecture for this work, achieving maximum training and validation accuracies of 97% for pistachio adulteration with peas. This research will yield sophisticated AI software that facilitates the seamless online classification of adulterated pistachio in the future.