Simulating internal watershed processes using multiple SWAT models

Highlights

• Internal watershed processes depend on model inputs, parameters, and structure

• These model variations were captured through use of multiple models

• Edge of field data was used to evaluate internal watershed process simulation

• Calibration at watershed outlet did not ensure accurate simulation upstream

• Model structure limited accurate internal partitioning of P transport

Abstract

The need for effective water quality models to help guide management and policy, and extend monitoring information, is at the forefront of recent discussions related to watershed management. These models are often calibrated and validated at the basin outlet, which ensures that models are capable of evaluating basin scale hydrology and water quality. However, there is a need to understand where these models succeed or fail with respect to internal process representation, as these watershed-scale models are used to inform management practices and mitigation strategies upstream. We evaluated an ensemble of models—each calibrated to in-stream observations at the basin outlet—against discharge and nutrient observations at the farm field scale to determine the extent to which these models capture field-scale dynamics. While all models performed well at the watershed outlet, upstream performance varied. Models tended to over-predict discharge through surface runoff and subsurface drainage, while under-predicting phosphorus loading through subsurface drainage and nitrogen loading through surface runoff. Our study suggests that while models may be applied to predict impacts of management at the basin scale, care should be taken in applying the models to evaluate field-scale management and processes in the absence of data that can be incorporated at that scale, even with the use of multiple models.

Introduction

Shifts in land use and management have transformed many landscapes, altering water resources and raising environmental concerns among managers and other stakeholders (Guswa et al., 2014). Understanding and predicting the impacts of these changes is key to developing effective mitigation strategies that balance development and environmental goals. Policy- and decision-making often rely on computer model simulations, raising concerns around appropriate model use and prediction confidence. Significant progress in model error analysis and uncertainty estimation has been made; however, models considered to be accurate can have inaccurate representations of dominant processes in a system (Hrachowitz and Clark, 2017). Models varying in structure and parameterization can effectively reproduce similar system behavior—the equifinality concept—allowing for multiple possible internal representations of a system (Beven, 2006). Capturing this process-level information results in more robust hydrologic and land management models and is key for effective decision-making.

Distributed parameter models have become a common tool for hydrologic analysis and watershed management planning. With greater access to spatially variable and fine resolution datasets, the representation of natural systems within these models can be vastly improved (Vieth et al., 2008; Daggupati et al., 2015). The Soil and Water Assessment Tool (SWAT) is a highly parameterized process-based model developed for watershed scale assessment of climate, land management, and land-use change (Neitsch et al., 2011). While originally intended for use on ungauged basins, modelers generally improve SWAT model performance through calibration and validation to streamflow, and potentially water quality, at a single downstream outlet location. To the extent that models are only assessed at the watershed outlet, they can fail to take into account key intra-watershed processes—internal processes which govern global responses—and can thereby lead to successful models that lack realistic system representation (Yen et al., 2014; Daggupati et al., 2015). Yet capturing these processes is important as most management implementation and improvement strategies in agricultural regions occur at the field scale (Diebel et al., 2008; Muenich et al., 2017).

Data capable of validating these internal processes at the necessary scales are rarely available, especially in large watersheds. This use of smaller scale data is important as studies have shown that, while sometimes difficult to incorporate and calibrate to, the use of direct field-scale data can significantly impact field-scale model performance (Merriman et al., 2018; Muenich et al., 2017; Kalcic et al., 2015; Daggupati et al., 2015). Two important sources of field-scale data that have become more available and have been used in upstream validation are remote-sensed data and edge of field (EOF) monitoring. The use of remotely sensed data, capable of effectively describing the internal distribution of key watershed characteristics, such as soil moisture (Rajib et al., 2016) and vegetation (Ma et al., 2019; Rajib et al., 2020), has greatly improved SWAT calibration and performance. EOF monitoring provides valuable site-specific information on management practices, as well as key nutrient transport budgets (Pease et al., 2017; Hanrahan et al., 2019), which can inform soft validation for upland performance. As their availability grows, these data present an opportunity to add to the understanding and refinement of upstream representation and prediction accuracy of hydrologic models.

A lack of data availability, compounded with model framework uncertainties and computational constraints, may prohibit thorough uncertainty characterization of deterministic models such as SWAT. The use of multiple models that vary in model inputs, parameterization, and structure can capture some of the variability that is lost using a single deterministic model, and therefore play a role in uncertainty analysis (Hrachowitz and Clark, 2017). Such ensemble modeling has been used in a variety of SWAT modeling efforts to evaluate watershed water quality and policy-relevant management (Evenson et al., 2020; Martin et al., 2019; Scavia et al., 2017). Model ensembles can help capture uncertainties in model structure, parameterization, input data, and assumptions, thus enabling the exploration of the impacts of these facets on model performance.

The aim of this study was to assess the extent to which watershed hydrologic models, calibrated and validated at the watershed outlet, can capture field-scale dynamics upstream. This was done by developing the latest version of the Maumee River Watershed (MRW) SWAT model, Apostel, and comparing it against EOF monitoring data and two previous MRW SWAT model iterations, Kalcic et al. (2016) and Kujawa et al. (2020). While the Kalcic and Kujawa models varied with respect to management inputs, structure, and parameterization, both achieved acceptable Moriasi et al., 2007, Moriasi et al., 2015 performance metrics at the watershed outlet (Kalcic et al., 2016; Kujawa et al., 2020). Our objective was to assess the ability of the model ensemble to capture upstream field level discharges and nutrient loadings. The results will be discussed in the context of critical differences between models in terms of structure and inputs.