Understanding robustness in multiscale nutrient abatement: Probabilistic simulation-optimization using Bayesian network emulators

Abstract

Ecosystem management in the face of uncertain disturbances has triggered increasing practices of resilience thinking. A multiscale probabilistic simulation-optimization framework is developed based on the nested nature of watersheds to inform decision robustness for Best Management Practices (BMPs). We presented a novel approach using hybrid Bayesian Networks (BNs) as interpretable and probabilistic emulators of process-based models. The hybrid BNs established at the scale of Hydrologic Response Units (HRUs) are embedded into simulation-optimization, whereby we analyze the cost-effectiveness-robustness of candidate BMP strategies at the subbasin scale. The optimal strategy is identified in compliance with water quality standards using watershed-scale integer programming. We apply the approaches in a typical intensively cultivated plateau watershed adjacent to Lake Dianchi, one of the three most eutrophic lakes in China. Our findings suggest that the hybrid BNs, incorporating both quantitative and qualitative information, are reliable emulators of the Soil and Water Assessment Tool (SWAT) in capturing critical pathways of diffuse phosphorus. Tradeoffs among cost, effectiveness, and robustness follow the law of diminishing marginal benefits. The optimum BMP strategies vary with policymakers’ preference toward robustness levels. Our findings indicate that robustness should be accounted for as an additional decision attribute besides costs and pollution mitigation. The benefits of the modeling framework are to (i) reduce over 99% computation complexity and support efficient decision-making under multifaceted uncertainties; (ii) improve interpretability and reliability of machine learning emulators; and (iii) inform policymakers of robustness with the probability of water quality restoration success.