Use of Unsupervised Machine Learning for Agricultural Supply Chain Data Labeling

The heterogeneous data produced in agricultural supply chains can be divided into three main systems: (i) product identification and traceability, related to identifying production batches and locations of the product throughout the supply chain; (ii) environmental monitoring, considering environmental variables in production, storage and transportation; and (iii) processes monitoring, related to the data describing the production processes and inputs used. Data labeling on the different systems can improve decision-making, traceability, and coordination in the chains. Nevertheless, this is a labor-intensive task. The objective of this Chapter was to evaluate if unsupervised machine learning techniques could be used to identify patterns in the data, clusters of data, and generate labels for an unlabeled agricultural supply chain dataset. A dataset was generated through merging seven datasets that contained information from the three systems, and the k-means and self-organizing maps (SOM) models were evaluated on clustering the data and generating labels. The use of principal component analysis (PCA) was also evaluated together with the k-means model. Several supervised and unsupervised learning metrics were evaluated. The SOM model with the Gaussian neighborhood function provided the best results, with an F1-score of 0.91 and a more defined clusters map. A series of recommendations for the use of unsupervised learning techniques on supply chain data are discussed. The methodology used in this Chapter can be implemented on other supply chains and unsupervised machine learning research. Future work is related to improving the dataset and implementing other clustering models and dimensionality reduction techniques.