News

In this proof of concept study (open access) the authors used machine learning to build a classification model for saffrom authentication using mid infrared spectroscopy (MIR).  MIR is a portable technique.

111 authentic saffron samples (2023 harvest) were directly sourced from farmers in 6 different regions of Iran.

 Adulterant were prepared in-house by blending 1 – 30% of saffron style, safflower, madder, or calendula.

After grinding  into a uniform powder and then sieved in accordance with ISO 3632 standard, the samples were washed, dried, ground and then extracted into ultrapure water.  Extracts were filtered, then analysed by a solvent MIR spectrometer.

The authors report that data-driven soft independent modeling of class analogy (DD-SIMCA) successfully differentiated between authentic and adulterated samples, achieving 100 % sensitivity and specificity. PLS-DA and RSDE were then employed to identify the type and level of adulterants, with RSDE clearly outperforming PLS-DA, achieving accuracy above 94.0 %, as compared to PLS-DA's accuracy of over 90.0 %. They were also able to differentiate between the 6 Iranian growing regions. The authors do not report if they challenged or validated their model with samples independent of the reference set.

In conclusion, they conclude that the combination of solvent-based MIR spectroscopy and modern chemometric techniques shows great potential as a reliable tool for saffron quality control at the point of need.

Photo by Vera De on Unsplash

In this study (purchase required) a machine vision system was used to capture the images of saffron samples at different safflower mixture proportions. Then three feature extraction algorithms - gray level co-occurrence matrix, gray-level run-length matrix, and Local Binary Pattern -  were applied to extract the textural features of data. Discriminant Analysis, Support Vector Machine, and Artificial Neural Network algorithms as supervised classification models were applied to classify datasets.

The models were applied for 3 class and 6 class datasets to explore classification ability. The best outcome for the 6-class dataset was with the Support Vector Machine model and with all features with an accuracy of 80 %. For 3 class datasets, Discriminant Analysis model had the best result with all features and with the accuracy of 97.78 %.

To explore the statistical importance of different features, two Minimum Redundancy Maximum Relevance and Chi-Square Test algorithms were applied. For the gray level co-occurrence matrix extracted features, Chi-Square Test algorithm with 10 features had the best accuracy with a test accuracy of 76.94 %.

The authors conclude that the proposed approach could be utilized in designing a system for checking saffron authenticity at a business-to-business point of sale..