Improving the stability of herbs through algorithms and machine learning

A predictive model for the stability of dried herbs based on production parameters and end product analyses. Due to their low altitude and proximity to the ground, herbs are susceptible to contamination with microorganisms. In addition, they are exposed to numerous environmental influences, which leads to strong quality fluctuations and high further processing costs.

The Austrian Mountain Herb Cooperative, a cooperative of innovative farmers in the Mühlviertel region of Upper Austria for the cultivation and sale of herbs from controlled organic farming, ventured to try to get a better handle on these quality fluctuations with the help of applied statistics and machine learning as part of a project with FFoQSI and the Bioinformatics Research Group of the Upper Austrian University of Applied Sciences in Hagenberg.
The cooperative, a cooperative of innovative farmers in the Upper Austrian Mühlviertel region for the cultivation and sale of herbs from controlled organic farming, dared to try to get a better grip on these quality fluctuations with the help of applied statistics and machine learning in a project with FFoQSI and the Bioinformatics Research Group of the Upper Austrian University of Applied Sciences in Hagenberg. The basis for this was the
and provided by the approximately 60 members of the cooperative. On the one hand, there were more than 100 parameters on the cultivation, harvesting and drying of each batch, such as conditions during planting, type and number of soil cultivation steps, type and applications of fertilisers, conditions during harvesting and the drying parameters; on the other hand, each batch was also examined in a laboratory for microbial contamination, e.g. with yeasts, moulds or pathogens such as salmonella. All these data were collected for each batch of each raw material. In order to predict the risk of microbial contamination in future batches, various machine learning algorithms were used, such as random forests, gradient boosting trees, artificial neural networks or symbolic regression. In addition, applied statistics and hypothesis testing were used to identify the most relevant parameters for high microbial contamination. Thus, farmers can be given concrete recommendations for action to reduce microbial contamination. All these aspects are communicated via a web application that serves both to collect new data and to present the evaluations to the farmers. In the course of its development, the application was continuously adapted to the needs of the farmers via several feedback meetings.

Effects and impacts
The information system developed in this project creates greater transparency for both the cooperative and the member farms through improved monitoring of the data. By analysing large data sets from different sources with machine learning, it has been possible to identify correlation and prediction models. This leads to a better understanding of the processes, a reduced microbial load and thus to an increase in product safety and storability of the herbs produced. The forecasting system is now to be further refined and improved by the end of 2024.

Scientific coordination: FH-Prof. PD DI Dr. Stephan Winkler, Head of Research Group, Bioinformatics in Hagenberg, University of Applied Sciences FH Upper Austria
Partner: Österreichische Bergkräutergenossenschaft eGen (ca. 60 member companies), Austria