Date of Award

Winter 2019

Project Type


Program or Major

Natural Resources

Degree Name

Master of Science

First Advisor

Shadi Atallah

Second Advisor

Marek Petrik

Third Advisor

Jeffrey Garnas


A pressing challenge of modern agriculture is to develop means of decreasing the negative impacts of pesticides while maintaining low pest pressure and high crop yield. Certain crop varieties, especially wild relatives of domesticated crops, provide pest regulation ecosystem services through chemical defense mechanisms. Benefits from these ecosystem service can be realized by intercropping cash crops with repellent wild varieties to reduce pest pressure. An opportunity cost exists, however, which consists of lower yield and market value. Such is the case of heirloom apple varieties that are more resistant to the codling moth but have a lower market value compared to commercial apples such as Red Delicious and Gala. In this thesis, I first develop a model to identify the bioeconomically optimal intercropping level of commercial and wild varieties with the purpose of pest management in the specific case of the codling moth. Second, I develop a model that uses a machine learning technique to determine pesticide application policies for the multi-variety orchard, where the solution is robust to model and data uncertainty.

Model 1 is a tree-level, spatially-explicit, bioeconomic simulation model. In the baseline case, we find that the bioeconomically optimal variety mix consists of 20% cider variety and 80% commercial variety. We analyze the sensitivity of the optimal mix to the market price difference of the two apple varieties and find that the optimal proportion of cider decreases linearly and that 100% commercial variety is optimal if the price difference is greater than $0.3/lb. We consider eight different spatial configurations for the intercropping, in addition to the baseline random spatial intercropping and find that the diagonal configuration yields the highest net present value and requires the lowest amount of cider intercropping (4%). Random spatial intercropping, in contrast, ranks seventh and has the second-highest optimal proportion of cider (30%). We use the certainty equivalent measure to determine how the optimal mix changes for a grower who has a moderate level of risk aversion, where production risk is driven by the effect of temperature on codling moth infestation over the years. The optimal cider variety percentage for a moderately risk-averse grower increases to 38% compared to the baseline case of 20% of a risk-neutral grower. We also document the risk-reducing effect of apple agrobiodiversity by characterizing how the risk premium decreases with increasing proportions of cider.

In Model 2, we determine the robust optimal pesticide application threshold, given an infested multi-variety orchard consisting of the optimal proportion of cider varieties, arranged in a random spatial configuration. We use historical degree-day (DD) data and associated established DD threshold-based spray recommendations to add pesticide application features to our Model 1 and then use it as a simulator to generate data on infestation and damage level over time. We then use Reinforcement Learning (RL) to find the robust optimal pesticide application threshold around 1,000 insects over the entire orchard. The model solution shows a greater degree of sensitivity to pesticide application costs compared to the pest growth rate, indicating the importance of addressing the data uncertainty of these parameters.