A data- driven, machine learning framework for optimal pest management in cotton

Despite the significant effects of agricultural pest management on crop yield, profit, environmental quality, and sustainability, farmers oftentimes lack datadriven decision support to help optimize pest management strategies. To address this need, we curated a comprehensive data set that consists of pest, pest management, and yield information from 1498 commercial cotton crops in California’s San Joaquin Valley between 1997 and 2008. Using this data set, we built a Markov decision process model to identify the optimal management policy of a key cotton pest, Lygus hesperus, that balances the tradeoff between yield loss and the cost of pesticide applications. Our results show that pesticide applications targeting L. hesperus are only economically optimal during the first 2 weeks of June, and pesticide applications were associated with increased risk of an unprofitable harvest. About 46% of the observations in our data set involved at least one pesticide application outside of this optimal window, demonstrating the need for a datadriven approach to crop management. Sensitivity analyses on parameter perturbations and reduced data set sizes suggest that our methodology provides a robust policymaking tool, even in noisy data sets.