Collaborative Research: MRA: Advancing process understanding of lake water quality to macrosystem scales with knowledge-guided machine learning

尽管人类活动对湖泊的影响越来越大，但关于大陆尺度湖泊水质的信息却非常稀少。此外，对于气候和土地利用等关键驱动因素的大范围变化如何控制大陆范围内的水质，我们还只有初步的了解。因此，理解基于几个相对研究较好的湖泊的生态学知识如何适用于美国大陆是一个挑战，那里的数据有限，涉及的湖泊有数千个。由于湖泊数量众多，水质问题复杂，机器学习可能会被证明是有用的。然而，机器学习的最新进展在商业应用中显示出巨大的成功，但尚未完全应用于自然系统中的问题，如湖泊水质，部分原因是数据量较低。此外，基础生态学研究的一个基本目标是机械地理解世界的运作方式，这是许多机器学习方法所缺少的目标。该项目开发了生态知识引导的机器学习(Eco-KGML)作为一个框架，以利用生态理解和机器学习的力量对美国各地的湖泊水质进行建模。Eco-KGML提高了水质预测的准确性，并促进了对水质过程的新知识的发现。为了扩大这项工作的影响，该项目通过一项培训计划，支持妇女和代表不足的少数族裔参与STEM(科学、技术、工程和数学)，该培训计划由从历史上被排除在外的群体招募的本科生队列组成，他们每年夏天都会参与Eco-KGML研究项目。该计划提供真实的研究体验，在学年期间演变为个人研究项目，并在支持性环境中让学生参与跨学科、跨机构、协作的科学。该项目还通过制作和传播向学生介绍生态-KGML概念的互动软件模块来改进STEM教育。通过与美国联邦机构合作伙伴和国家生态观测网络(NEON)的合作者合作，该项目的更广泛影响超出了参与大学的范围，这些合作伙伴向机构和霓虹灯优先事项提供信息并提供反馈。该项目开发了生态知识引导的机器学习(Eco-KGML)，作为在大系统尺度上模拟湖泊和水库水质(WQ)动态的概念框架。ECO-KGML使用基于动态过程的模型和ML模型的混合组合，在地理上广泛的WQ数据的帮助下，将WQ过程从经过充分研究的湖泊扩展到整个美国的宏观系统尺度。该项目侧重于水透明度、浮游植物生物量和低磁缺氧的具体WQ指标，以解决以下问题：控制水质的主导过程是什么，它们如何在空间和时间上变化？气候、土地利用和生态系统记忆如何相互作用，影响从局部到大系统尺度的水质动态？湖泊水质的大的空间和长期的变化模式是什么？在解决这些问题时，WQ湖的Eco-KGML模型被赋予了一条新的研究路线，该模型不仅旨在改善WQ变量的预测性能，而且还可以在一系列时空尺度上发现关于WQ过程的新知识。利用最大似然方法，探索了在给定WQ观测的情况下，以计算高效和可推广的方式估计湖泊过程参数的新研究。基于ML的湖泊WQ模型能够发现每个湖泊的WQ变量之间的新关系，并提取相关的时间滞后。通过对模块化组合学习(MCL)的新研究，开发了Eco-KGML模型来识别在给定的湖泊中哪些WQ过程是主导的，以及它们如何相互作用来影响整体WQ动态。此外，Eco-KGML模型学习并区分单个湖泊特有的过程，以及根据其生态特征概括不同类型湖泊的过程。这种对科学知识和数据的灵活和全面的使用使我们能够研究湖泊及其驱动因素之间的尺度依赖关系，同时为湖泊在多个时间和空间尺度上提供更可靠的预测。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Despite the growing influence of human activities on lakes, there is remarkably sparse information on lake water quality at continental scales. Moreover, we have only a nascent understanding of how broadscale changes in key drivers, such as climate and land use, control water quality at continental scales. Thus, it is a challenge to understand how ecological knowledge, based on a few relatively well-studied lakes, applies to the continental U.S., where data are limited for 1000s of lakes. Because of the large number of lakes and the complexity of the water quality problem, machine learning may prove useful. However, recent advances in machine learning that have shown great success in commercial applications have yet to be fully applied to problems in natural systems, such as lake water quality, in part because of lower data volumes. In addition, a fundamental goal of basic ecological research is mechanistic understanding of the way the world works, a goal missing in many machine learning approaches. This project develops ecology-knowledge guided machine learning (Eco-KGML) as a framework for leveraging the power of both ecological understanding and machine learning in modeling lake water quality across the U.S. Eco-KGML improves the accuracy of water quality predictions and advances the discovery of new knowledge about water quality processes. To broaden the impacts of this work, the project supports participation of women and underrepresented minorities in STEM (science, technology, engineering, and math) through a training program consisting of cohorts of undergraduate students, recruited from historically-excluded groups, who work on Eco-KGML research projects each summer. This program provides authentic research experiences that evolve into individual research projects during the academic year and engage students in cross-disciplinary, cross-institutional, collaborative science in a supportive environment. This project also improves STEM education through production and dissemination of an interactive software module that introduces students to Eco-KGML concepts. The broader impact of this project extends beyond the participating universities through collaborations with U.S. federal agency partners and collaborators from the National Ecological Observatory Network (NEON) that inform, and feed back to, agency and NEON priorities. This project develops ecology-knowledge guided machine learning (Eco-KGML) as a conceptual framework for modeling lake and reservoir water quality (WQ) dynamics at macrosystem scales. Eco-KGML uses hybrid combinations of dynamical process-based models and ML models to scale WQ processes from well-studied lakes to macrosystem-scales across the U.S with the help of geographically extensive WQ data. This project focuses on the specific WQ metrics of water clarity, phytoplankton biomass, and hypolimnetic anoxia, in addressing the questions: What are the dominant processes governing water quality and how do they vary across space and time? How do climate, land use, and ecosystem memory interact to affect water quality dynamics from local to macrosystem-scales? What are the broad spatial and long-term patterns of change in lake water quality? In addressing these questions, a new line of research is enabled in Eco-KGML models for lake WQ, which are not only aimed at improving predictive performance of WQ variables but can also lead to discovery of new knowledge about WQ processes at a range of spatio-temporal scales. Novel research in estimating process parameters of a lake, given its WQ observations, in a computationally efficient and generalizable manner is explored using ML methods. The ML-based models for lake WQ enable the discovery of new relationships among WQ variables at every lake, along with extracting relevant time lags. Through novel research in modular compositional learning (MCL), Eco-KGML models are developed to identify which WQ processes are dominant at a given lake and how they interact to influence overall WQ dynamics. Moreover, the Eco-KGML models learn and distinguish processes specific to a single lake from those that generalize across types of lakes according to its ecological characteristics. This flexible and comprehensive use of both scientific knowledge and data enable the study of scale-dependent relationships between lakes and their drivers while providing more robust predictions for lakes across multiple temporal and spatial scales.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.