Using Machine Learning to Investigate Weather Impacts on Crop Yield for Probabilistic, Climate-Informed Predictive Crop Breeding

As the world’s population continues to grow, the importance of food security will only continue to increase, with some estimates requiring a global increase in food production between 59 and 98% by 2050. The agriculture industry is also a crucial economic driver, with the industry supporting nearly 25% of the state of North Dakota’s workforce, and with a total valuation of $9 billion. The economic importance of this industry and the challenges of future growth prompt a need for data driven decision making into the future, particularly with predictive crop breeding adaptive to a changing climate. Using the North American Regional Reanalysis as well as data from the National Agricultural Statistics Survey, we implement a machine learning based methodology to determine what atmospheric variables are important for predicting yields for a variety of crops across the state of North Dakota. Non-linear relationships between atmospheric variables and crop yield encourage the use of machine learning to diagnose the importance of each variable. We utilize a random forest ensemble learning algorithm to determine the importance of each atmospheric variable. Feature importance is assessed using the permutation method, which measures the decrease in model performance when a feature's values are randomly shuffled, indicating its relevance. The goal is to identify what variables are most important for predicting yield as well as their timing in relation to the growing season. With variable importance defined, changes and predictability of these changes in future climates can be used to assist predictive crop breeding programs.