Bilevel optimization of conservation practices for agricultural production

Uncertainty surrounding the sources of nonpoint pollution and producer response to prospective policy incentives complicates the nonpoint policy problem. To explain, neither policy-makers nor producers know the exact effect of current and alternative farming practices on the contributions of specific cropped fields to nutrient pollution. Spatial heterogeneity of the production technology and environmental damage of runoff also precludes the formulation of an analytic solution, so that producer response to prospective policies is unknown a priori to the policy maker. To address the complexity arising from point source uncertainty and spatial heterogeneity, we draw on recent computational advances to reformulate the classic nonpoint source pollution problem as a multi-objective bilevel optimization problem, employing genetic algorithm (GA) solution methods. This computational framework explicitly accounts for the nested nature of farm-level management decisions in response to prospective agri-environmental policy incentives, and spatial heterogeneity of both production and pollution effects. Our application considers the optimal spatial targeting of multiple management practices in the Iowa Raccoon River Watershed, an intensive corn and soybean production region of the Upper Mississippi River Basin (UMRB). Consistent with theory and previous empirical results, we find that combining multiple management practices, versus relying on single or one-size policies, lowers the total cost for a given level of nitrogen reduction. But we also find overall limited potential nitrogen reduction via implementing these practices on working land, suggesting the continued need for land retirement in meeting current nonpoint policy goals for the UMRB. We believe that the main contribution of this study lies in the novelty of the bilevel approach, which explicitly accounts for feedbacks between policy makers and agricultural producers, while the associated GA computational methods allow for better handling of large scale and complex spatial heterogeneity over the agricultural watershed.