Abstract/Summary

Grain protein content is the most important indicator of wheat quality; it is affected by environmental conditions and agronomic practices. Thus, predictions at an early stage before harvest are crucial for farmers to decide their agronomic practices. This paper describes the development of a machine learning approach (MLA) based on the Bayesian networks (BNs) model to predict grain protein content using soil, topographic and yield data. The model has been developed using a Bayesian belief network software, categorising each node within each field based on the data available for a given field. The conditional interdependencies of these variables were learned using 75% of the data and then applied to 25% of the data to test the model. Grain protein content predictions were based on the probability of 50% chance of observing. The correlation between the predicted protein content and actual protein content was 0.40 and 0.48 for the German and UK test fields respectively.