Date of Award

5-14-2023

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Earth Sciences

Advisor(s)

Christa Kelleher

Keywords

Hydrology, Machine Learning, Modeling, Water Quality

Abstract

Models are universal tools that can be used to improve our understanding of natural system functioning and how these systems respond to perturbations (i.e., changing conditions). Due to the growth of computing power over the last decades, increasingly detailed models capable of working at finer resolutions and larger scales are expected to play a key role in advancing understanding Earth system sciences (Wood et al., 2011; Bierkens et al., 2014; Rauser et al., 2016). Such models are often used for weather forecasts, flood projection, drought severity assessment, among countless other applications including atmospheric circulations (Cotton et al., 2014), biochemical processes (Soetaert et al., 2000), habitat distributions (Guisan et al., 2000), and tsunami impacts (Gelfenbaum et al., 2011). Global climate models simulate interactions between multiple climate drivers, such as the atmosphere, oceans, land surface, and ice, producing future climate predictions, while providing an insight into past, current, and future dynamics (Schewe et., 2014). Results of such systems modeling exercises are often used to influence policy decisions; hence, making accurate projections on changing climate and understanding system response to these changes can be critical.

Complex model frameworks often require new approaches and advanced physical process understanding to achieve hyper-resolution prediction capability. Consequentially, computer models have become a popular tool to evaluate complex physical processes. Particularly in hydrologic sciences, as we are not able to measure everything we would like to know about hydrologic systems, models are used as a mean to extrapolate from those measurements in space and time (Beven, 2012). Although models can be physically-based, conceptual, or statistical representations of a given system, the main goal of any model is to represent the system as accurately as possible, in order to ensure that predictions and information obtained from the model are realistic (Oreskes et al., 1994; Sarrazin et al., 2016; Pianosi et al., 2016). While physically-based models represent a physical construct with characteristics that resemble the physical characteristics of the modeled system (Sparling, 2016), conceptual and mathematical models are more interrelated and abstract. Although conceptual models can be considered idealized representations of natural system (Klemes, 1986; Grayson et al., 1992; Wagener et al., 2005; Kirchner, 2006; Beven, 2007; Gupta et al., 2012; Hrachowitz et al., 2014), they are often backed by mathematical expressions (Wagener et al., 2003), while statistical models are solely based on empirical relationships (Breiman, 2001). Regardless of the model type, all models can provide valuable insight into the natural system and further advance process understanding.

We used conceptual and data-driven modeling to improve our process-based understanding in three different scenarios. First, we used time-varying sensitivity analysis combined with a conceptual rainfall-runoff model to explore how different hydroclimatic conditions impact model parameter importance. We found that sensitivity index variations through time could be organized into five unique patterns. These groupings were spatially organized, revealing the role of air temperature and precipitation in shaping parameter importance. Second, we used statistical analyses and empirical modeling to evaluate how various temporal and spatial variables influence presence and concentrations of Compounds of Emerging Concern (CECs). Results showed that changing hydroclimatic conditions and land cover likely play an important role in shaping observed CEC concentrations. These identified relationships were then used to inform the development of a Machine Learning (ML) model capable of predicting daily pesticide concentrations. In addition, in depth model diagnostics was used to determine which ML techniques are best suited to study pesticides like atrazine, with implications for other CECs. Results from these studies further demonstrate the versatility of conceptual and data-driven models that can be used to study questions that are often difficult to answer without utilizing numerical methods.

Access

Open Access

Download

Share

COinS