A global sensitivity analysis approach for identifying critical sources of uncertainty in non-identifiable, spatially distributed environmental models: A holistic analysis applied to SWAT for input datasets and model parameters

https://doi.org/10.1016/j.envsoft.2020.104676Get rights and content

Highlights

  • A holistic sensitivity analysis, including model structure uncertainty, is conducted for a spatially distributed model.

  • A sensitivity analysis examines the relative importance of the uncertainty sources using a SWAT application.

  • The uncertainties in the stream network precision and certain SWAT parameters are critical.

Abstracts

Environmental models have a key role to play in understanding complex environmental phenomena in space and time. Although their inherent uncertainty and non-identifiability are being increasingly recognized with the development and application of various methods, a more holistic analysis of all sources of model uncertainty is warranted. This paper addresses sources of uncertainty from various types of input datasets and model parameters, including those related to model structure assumptions, using a Soil and Water Assessment Tool (SWAT) application for the Minjiang River watershed, China. The holistic uncertainty sources in the SWAT application are summarized, and a sensitivity analysis (SA) is applied to examine the relative importance of the uncertainty sources influencing average streamflow and the load of nitrate. The analysis reveals that uncertainties related to the stream network precision and certain SWAT parameters are the most critical factors. Furthermore, building upon our SA framework to consider uncertainty sources more holistically would provide a good starting point for subsequent SA of spatially distributed environmental models in general.

Introduction

Deterministic environmental models offer useful methods for exploring problems, providing predictions, and supporting decisions that involve complex environmental phenomena evolving in space and time. It has become increasingly recognized, however, that these models are almost always non-identifiable (Guillaume et al., 2019), in the sense that the quantity and quality of data are insufficient to parameterize the models uniquely, and that the associated modelling must address a wide range of uncertainties, especially those related to predicting the impact of possible management actions. Accordingly, Uusitalo et al. (2015) review various methods that can be applied to evaluate uncertainty of deterministic model outputs. They cover expert assessment, model emulation, sensitivity analysis (SA) and use of multiple models, arguing that the best method for uncertainty evaluation is determined by the definition of a model, and the amount of available information. SA is an established approach to model assessment by quantitatively evaluating the change in model output(s) with respect to changes in model factors (usually parameters, forcing and other input data). A deterministic environmental model can be easily coupled with SA based on Monte Carlo type simulation (Farmer and Vogel, 2016). SA offers a quantitative evaluation for uncertainty in a model, rather than a qualitative evaluation which, for example, expert assessment does (Uusitalo et al., 2015), although qualitative evaluation could provide additional evidence to further support decision making and might be more efficient when a modeler is well-informed. In this paper we employ a global method of SA (Saltelli et al., 2008) as opposed to a local method. A global method allows one to compute the contribution to output sensitivity over a plausible range of factor values.

Global SA is undertaken here as a useful first step in addressing and understanding the criticality of uncertainty sources (see Norton, 2015 for an overview of methods, and Crosetto and Tarantola, 2001; Gan et al., 2014; Tong, 2015) by assessing those factors, and their plausible range of values, that dominate changes in model outputs. Understanding the criticalities can then be valuable for undertaking an uncertainty analysis in the traditional probabilistic sense, for example, where one assumes prior distributions of parameters and finds posterior distributions based on some measure of likelihood.

As indicated in the literature review of Section 2, previous approaches to SA tend to address only a modest set of the sources of uncertainty for spatially distributed environmental models. The aim of this article is to illustrate how a more holistic SA approach to spatially distributed environmental models can be used to identify their critical sources of uncertainty, which would subsequently allow focus on them for more specific uncertainty analysis and even its reduction. The Soil and Water Assessment Tool (SWAT) (Neitsch et al., 2011) is used as an archetypal example as it is the most frequently used water quantity-quality model (Fu et al., 2019; Ray, 2018). It is also a typical example of the type of uncertainties that need to be considered in a spatially distributed environmental model. The global SA approach undertaken is more complete than previous SA studies on spatially distributed environmental models in the sense that: it attempts to address model structure uncertainty in combination with the usual model parameter and data uncertainties; and examines impact on average streamflow as well as water quality outputs. Thus, this article addresses the uncertainty of model structure input parameters related to the submodels of a SWAT application [i.e., watershed delineation and hydrological response unit (HRU) characteristics]. It also explores the measurement uncertainty of the digital elevation model (DEM) (i.e., its vertical accuracy), and the uncertainty of boundaries of classes in land-use-land-cover (LULC) and soil datasets (Crosetto and Tarantola, 2001; Goodchild and Guoqing, 1992) because they have a profound effect on watershed delineation (Oksanen and Sarjakoski, 2005; Wu et al., 2008) and HRU creation, respectively. Another aspect investigated is the impact of measurement errors in meteorological information on model outputs. Finally, the SA is also applied to the model parameters investigated as has been the primary focus in the past (e.g., Setegn et al., 2010; Wu and Liu, 2012; Yang et al., 2018; Zhao et al., 2018), some of which relate to model structure assumptions. Clearly, while it is not possible to investigate all model structure assumptions as these can be innumerable in an environmental modelling exercise, it is acknowledged that expert assessment is a key, complementary and qualitative method that can be used to justify other model structure assumptions in relation to uncertainty (O'Hagan, 2012; Uusitalo et al., 2015). Our SA approach could therefore be followed up with a formal uncertainty analysis to help in setting the range of prior distributions and fixing unimportant factors, which is especially important when sampling needs to be limited because of high computational demands of the environmental model.

The remainder of this article is organized as follows. Section 2 contains a literature review and key qualitative findings of previous SA studies on the SWAT model. Section 3 briefly explains SWAT and the extended Fourier Amplitude Sensitivity Test (FAST) method which is the global SA method undertaken here. Then, in Section 4, the detailed holistic SA process is reported with our SWAT application to the Minjiang River watershed in Sichuan, China. This analysis follows a general process of SA (Cheng et al., 2014; Gan et al., 2014): identifying uncertainty sources associated with submodels in SWAT, and propagating the uncertainty from the identified source. Results of analyzing the uncertainty sources using SA appear in Section 5. Specifically, this article identifies uncertainty sources related to model structure input parameters and datasets, as well as general model parameters. Then, uncertainty propagation methods are utilized with respect to the corresponding uncertainty sources. For spatial input datasets, the measurement uncertainty of a DEM is propagated using a sequential Gaussian simulation to represent spatially autocorrelated uncertainty (Goovaerts, 1997; Pebesma, 2004), and the boundary uncertainties of LULC and soil datasets are simulated by adopting the epsilon band approach (Crosetto and Tarantola, 2001; Shi, 1998). The SA evaluates the relative importance of the uncertainty sources in the average streamflow (FLOW) and loads of nitrate (NO3). This article concludes in Section 6 with a discussion on future work that could profitably be conducted in relation to our analysis.

Section snippets

Previous SA studies on SWAT

There has been too little attention given to sensitivity analysis of spatially distributed models that focuses on a wide range of uncertainties, especially the influence of model structure assumptions. In illustration, this section explores limitations of previous SA studies for the water quantity-quality model known as SWAT, partly because it provides a typical example of the limitations of previous sensitivity studies with respect to a spatially distributed environmental model. SWAT requires

SWAT

SWAT is a watershed model that was developed by the Agricultural Research Service of the U.S. Department of Agriculture (USDA) (Neitsch et al., 2011). SWAT has been extensively applied in various sectors, including water management, hydrology, climate change, land use impact, and pollution, to predict the environmental impacts (e.g., land use and climate change) on water quantity and quality. To implement SA for SWAT applications (Bastin et al., 2013), this paper uses a tool for automated SWAT

Study area and datasets

This study illustrates a more holistic SA approach to spatially distributed environmental models with a SWAT example in a subset of the Minjiang River watershed. The Minjiang River watershed is located in the upper part of the Yangtze River basin, and the subset of the watershed in this analysis covers approximately 12,893 km2 (Fig. 1). The Minjiang River is the largest tributary of the Upper Yangtze River and supplies water to downstream regions for agriculture (e.g., Chengdu and its

Results of analyzing the uncertainty sources through SA

This analysis evaluates the impacts of the fifteen factors on variations in the two SWAT output estimates considered; i.e., FLOW and NO3 at the outlet of the Minjiang River watershed (refer to Fig. 1). Firstly, the impacts of the fifteen factors on FLOW are evaluated in terms of the main and total effect indices using the extended FAST (Fig. 4). As we discussed in Section 4.3, two different ranges of MinStream [i.e., the wide range (W-SA) between 5,000 and 15,000 and the narrow range (N-SA)

Conclusion

A wide range of sources of uncertainty and their relative importance were examined in the context of an integrated, spatially distributed environmental model based on a SWAT application in the Minjiang River watershed in Sichuan, China. This paper identified the strength of uncertainty sources with respect to watershed delineation, HRU creation, and SWAT execution, thereby addressing the uncertainty of model structure through uncertainties of the input parameters and datasets in these

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

We thank the OpenGMS team from Nanjing Normal University for their assistance with an automated SWAT preparation tool. This research was supported by the NSF for Excellent Young Scholars of China, Grant 41622108, the National Basic Research Program of China, Grant 2015CB954103, and the Priority Academic Program Development of Jiangsu Higher Education Institutions, Grant 164320H116.

References (82)

  • J. Norton

    An introduction to sensitivity assessment of simulation models

    Environ. Model. Software

    (2015)
  • A. O'Hagan

    Probabilistic uncertainty specification: overview, elaboration techniques and their application to a mechanistic model of carbon flux

    Environ. Model. Software

    (2012)
  • F. Pappenberger et al.

    Multi-method global sensitivity analysis of flood inundation models

    Adv. Water Resour.

    (2008)
  • E.J. Pebesma

    Multivariable geostatistics in S: the gstat package

    Comput. Geosci.

    (2004)
  • S. Razavi et al.

    A multi-method Generalized Global Sensitivity Matrix approach to accounting for the dynamical nature of earth and environmental systems models

    Environ. Model. Software

    (2019)
  • Z. Shen et al.

    Parameter uncertainty analysis of the non-point source pollution in the Daning River watershed of the three Gorges Reservoir region , China

    Sci. Total Environ.

    (2008)
  • Z.Y. Shen et al.

    A comprehensive study of the effect of GIS data on hydrology and non-point source pollution modeling

    Agric. Water Manag.

    (2013)
  • I.M. Sobol

    Global sensitivity indices for nonlinear mathematical models and their Monte Carlo estimates

    Math. Comput. Simulat.

    (2001)
  • M. Strauch et al.

    Using precipitation data ensemble for uncertainty analysis in SWAT streamflow simulation

    J. Hydrol

    (2012)
  • A. Tasdighi et al.

    A Bayesian total uncertainty analysis framework for assessment of management practices using watershed models

    Environ. Model. Software

    (2018)
  • A.J.A.M. Temme et al.

    Geostatistical simulation and error propagation in geomorphometry

    Dev. Soil Sci.

    (2009)
  • J. Teng et al.

    Flood inundation modelling: a review of methods, recent advances and uncertainty analysis

    Environ. Model. Software

    (2017)
  • L. Uusitalo et al.

    An overview of methods to evaluate uncertainty of deterministic models in decision support

    Environ. Model. Software

    (2015)
  • P.A. Vanrolleghem et al.

    Global sensitivity analysis for urban water quality modelling: terminology, convergence and comparison of different methods

    J. Hydrol

    (2015)
  • Y. Wu et al.

    Automating calibration, sensitivity and uncertainty analysis of complex models using the R package Flexible Modeling Environment (FME): SWAT as an example

    Environ. Model. Software

    (2012)
  • J. Yang et al.

    Uncertainty analysis of a semi-distributed hydrologic model based on a Gaussian Process emulator

    Environ. Model. Software

    (2018)
  • J. Yang et al.

    Hydrological modelling of the Chaohe basin in China: statistical model formulation and Bayesian inference

    J. Hydrol

    (2007)
  • F. Zhang et al.

    Design and development of a service-oriented wrapper system for sharing and reusing distributed geoanalysis models on the web

    Environ. Model. Software

    (2019)
  • K.C. Abbaspour

    SWAT-CUP: SWAT calibration and uncertainty programs - a user manual. Eawag: Swiss federal institute of aquatic science and technology

  • ANSI (American National Standards Institute)

    Spatial Data Transfer Standard (SDTS) - Part 1. Logical Specifications

    (1998)
  • J. Aouissi et al.

    Sensitivity analysis of SWAT model to the spatial rainfall distribution and watershed subdivision in streamflow simulations in the Mediterranean context : a case study in the Joumine watershed . Tunisia

  • A. Bárdossy et al.

    Influence of rainfall observation network on model calibration and application

    Hydrol. Earth Syst. Sci.

    (2008)
  • I. Chaubey et al.

    Effect of DEM data resolution on SWAT output

    Hydrol. Process.

    (2005)
  • T. Cheng et al.

    Handbook of Regional Science

    (2014)
  • J. Cho et al.

    Effect of spatial distribution of rainfall on temporal and spatial uncertainty of SWAT output

    Trans. ASABE (Am. Soc. Agric. Biol. Eng.)

    (2009)
  • N.R. Chrisman

    The error component in spatial data

  • G.J. Ciach

    Local random errors in tipping-bucket rain gauge measurements

    J. Atmos. Ocean. Technol.

    (2002)
  • M. Crosetto et al.

    Uncertainty and sensitivity analysis: tools for GIS-based model implementation

    Int. J. Geogr. Inf. Sci.

    (2001)
  • X. Cui et al.

    Impacts of Forest Changes on Hydrology : a Case Study of Large Watersheds in the Upper Reaches of Minjiang River Watershed in China 4279–4290

    (2012)
  • B. Dixon et al.

    Resample or not ?! Effects of resolution of DEMs in watershed modeling

    Hydrol. Process.

    (2009)
  • FAO/UNESCO

    The digitized soil map of the World and derived soil properties (version 3.5)

    FAO L. Water Digit. Media Ser

    (2003)
  • Cited by (42)

    View all citing articles on Scopus
    View full text