Bayesian calibration and benefits for environmental management

在水质评价和管理中，数学模型用于理解生态过程，预测水生生态系统动态，评估管理方案，并支持政策制定过程。环境模型包含由模型结构和参数造成的大量不确定性。模型不确定性的问题很重要，因为模型用于识别污染者，指导研究经费的使用，并确定具有重大社会和经济影响的管理战略。错误的模型输出和未能考虑到不确定性可能会产生误导性的结果，并在执行代价高昂的替代环境管理计划期间错误地分配有限的资源。本研究计划旨在通过将环境数学建模与贝叶斯分析相结合，解决对新型建模工具的迫切需求，从而有效地支持环境管理。基本假设是，贝叶斯方法提供了一种方便的方法，在这种方法中，过去的经验以及当前的生态信息构成了对未来生态系统反应更现实的预测的基础。检查贝叶斯技术的主要重点将放在湖泊富营养化模型上，但提议的框架可以很容易地转移到各种学科（例如，水文学、生态毒理学、空气污染）。加拿大退化最严重的三个大湖（安大略湖、伊利湖和温尼伯湖）将被用作案例研究。本研究的一些预期好处，如控制模型预测中的不确定性，优化监测程序的抽样设计，与适应性管理的政策实践保持一致（或“边做边学”），以及将模型输出表示为概率分布，对于利益相关者和政策制定者在制定可持续环境管理决策时特别有用。

In water quality assessment and management, mathematical models are used to understand ecological processes, to predict aquatic ecosystem dynamics, to evaluate management alternatives, and to support the policy making process. Environmental models involve substantial uncertainty contributed by both model structure and parameters. The question of model uncertainty is important because models are used to identify polluters, to direct the use of research dollars, and to determine management strategies that have considerable social and economic implications. Erroneous model outputs and failure to account for uncertainty could provide misleading results and misallocation of the limited resources during the costly implementation of alternative environmental management schemes. This research program aims to address the urgent need for novel modelling tools that can effectively support environmental management by combining environmental mathematical modelling with Bayesian analysis. The basic hypothesis is that the Bayesian approach provides a convenient methodology in which past experience along with present ecological information form the basis for more realistic predictions of future ecosystem response. The main focus for the examination of the Bayesian techniques will be placed on lake eutrophication models, but the proposed framework can be easily transferred to a wide variety of disciplines (e.g., hydrology, ecotoxicology, air pollution). Three of the most degraded Canadian large lakes (Ontario, Erie, and Winnipeg) will be used as case studies. Some of the anticipated benefits from this research, such as the control of uncertainty in model predictions, optimization of the sampling design of monitoring programs, alignment with the policy practice of adaptive management (or "learning by doing"), and expression of model outputs as probability distributions will be particularly useful for stakeholders and policy makers when making decisions for sustainable environmental management.