1. Introduction
Nutrient transport to water ways is of great concern to proper land management [1] [2]. Fertilizer application, usually a combination of nitrogen, phosphorus, and potassium, is one of the important inputs in crop production [3]. However, when nitrogen is overapplied or when the nitrogen use efficiency of a crop is low, the excess of nitrogen is transported to waterbodies or leached into groundwater, which can have far-reaching effects [1]. The effects include pollution such as water contamination and eutrophication of downstream sites, and nitrogen loss from the field and hence reduced nitrogen use efficiency of crop, and increased fertilizer costs to farmers [1] [4] [5]. Hydrologic and water quality models such as the Soil and Water Assessment Tool (SWAT) [6] and the Agricultural Policy/Environmental eXtender (APEX) [7] have been widely used to quantify the impacts of various management systems on water resources [8] [9] [10] [11] [12].
There are many studies comparing the impact of models [ [12] [13], etc.], soil [ [13] [14] [15] [16] [17], etc.] and weather [ [17] [18], etc.] data sources, evapotranspiration (ET) calculation methods [19], and digital elevation model (DEM) resolutions [ [20] [21], etc.] on model outputs, performance, and scenario results. However, based on the literature review, there are no studies that report the impact of the interfaces used to build models on model outputs and performance. Interfaces developed for hydrologic and water quality models are mainly used to pre-process data and create model input files for simulations [22] [23] [24]. However, different interfaces utilize different databases to derive model input. This can result in different default parameter values, which can lead to a different set of simulated outputs, conclusions, and recommendations. Model users choose an interface based on accessibility and ease of use and could benefit from a study that determines the potential impacts of the selected interface on model outcomes.
Models and interfaces utilize regional or national DEM, soils, and crop databases in order to provide default model input values for a given study area. Many studies present the advantages, disadvantages, and best methods of using databases in modeling [14] [15] [25] [26]. Some of these databases, such as soils and crop, may be modified by the model and interface developers based on the structure and possible quality assurance/quality control procedures. Although the developers give the users the option to modify the inputs, in most cases, model users don’t have measured data, especially at the watershed scale, to adjust the default parameter values. Therefore, many studies use the default values obtained from these databases without modification [19] [21] [26] [27] [28] [29] [30].
The APEX [7] [8] model is a watershed simulation model used to assess the impact of land management practices on water flow, sediment, and nutrients. APEX is a direct extension of Environmental Policy Integrated Climate Model (EPIC) [31]. There are five different interfaces used to process and build APEX model projects, including ArcAPEX [32], iAPEX [33], WinAPEX [34], APEX for Linux (https://epicapex.tamu.edu/model-executables/), and the Nutrient Tracking Tool (NTT) [35] [36]. Each of these interfaces have been used for different applications [15] [20] [29]. Monks et al. [15] used WinAPEX to compare the effects of different soil datasets on streamflow, surface runoff, and crop yields in Washington state, while Nelson et al. used ArcAPEX to build projects to compare the effect of the length of calibration period on hydrologic outputs [28] and examining the need for soft data in the calibration process [29]. Tadesse et al. [19] used NTT to compare the different evapotranspiration (ET) formulas available within the APEX model. One of the major structural differences between NTT and ArcAPEX interfaces is that ArcAPEX uses only the predominant soil for each subarea [37], while NTT assigns a maximum of three soils for every subarea, representing the most predominant soils in the area of interest [38]. Because model computation time takes place at the subarea level, this implies that a model built using NTT will require as much as three times the computation time to complete as one built by the ArcAPEX interface. However, one would hypothesize that although a model built using NTT requires more computation time, it should result in more realistic model outcomes because three soils for each subarea capture the variability better relative to the single soil used in ArcAPEX. However, none of the reported APEX literature presents the impact of the interface used on model outcomes. Therefore, the objectives of this study were to: 1) compare structure and input values of the ArcAPEX and NTT interfaces, and 2) determine the impact of the differences on simulated hydrology and water quality outputs, computation time, parameter sensitivity, and calibration performance.
2. Methods
2.1. Interface Input Structure
ArcAPEX is an ArcGIS-based user interface that incorporates soil data, topographic, land use, and a built-in APEX-Parameters database to simulate hydrologic and agricultural processes over a field to basin scale drainage area [32]. The NTT interface was developed to enable assessment of impacts of management practices and to facilitate water quality trading. It is a web-based interface with linkage to the APEX model [35].
The main APEX input files are CONTROL, PARM, Soils, and several management files (for operations, fertilizer, grazing, etc.). CONTROL and PARM files contain global parameters, meaning that these parameters are general and contain many coefficients used in different equations and the miscellaneous parameters used. The values of these parameters can be adjusted based on the crops, soils, and management practices representing the farming systems found in different regions of US [35] [38]. While the ArcAPEX and NTT interfaces utilize similar input files, there is a major structural difference with respect to the soil databases. Differences in soil databases include how soil properties are organized by layers for APEX and, more significantly, the number of soils for each subarea/or area of interest. In ArcAPEX, only the predominant soil is used for each subarea [37]. A dominant soil is assigned to each subarea from the list of soils in the study area (listed in the SOILCOM.DAT file). A file named filename.sol is used to describe each soil. The NTT interface allows users to verify, modify or delete soils copied from the SSURGO soil database and add or edit layers for the particular field selected in the field’s page. The NTT assigns a maximum of three soils for every subarea, representing the most predominant soils in the area of interest [38].
2.2. Interface Input Values
The values of the parameters in the Control, Parameter, and Soil files for the respective interfaces were determined after the model was built (see details below). Model building includes study area description, data sources, and model setup.
2.2.1. Study Area
Nelson et al. [29] provide a detailed description of the study area; thus only a summary is provided here. Rock Creek, located in northern Ohio, is a third order tributary of the Sandusky River (Figure 1), which flows north through the middle of Seneca County and drains into Lake Erie through Sandusky Bay [39] [40] [41].
Rock Creek watershed is approximately 7500 ha and has nineteen identified soil series, primarily from the Blount-Pewamo-Glynwood soil group [42]. These soils are moderately well drained to very poorly drained, and are located on slopes of 0% - 7%. Tile drainage occurs in ~90% of the agricultural fields, primarily in areas with 3% or less slope. The depth of tile drainage is approximately 0.9 m [43].
Seneca County’s climate is typical of the temperate mid-continent region. Rock Creek watershed is comprised of about 82% agricultural land, 13% forest land, and 6% urban land. Of the croplands, 50% are soybean, 30% are corn, and 20% are wheat [42]. Corn-soybean and corn-soybean-wheat are the most common crop rotations.
2.2.2. Data Sources
Three GIS data layers are required for the APEX model: digital elevation model (DEM), soils, and land use data. Sub-area parameters such as slope and slope length were calculated using a 30-m DEM obtained from the USGS
(http://viewer.nationalmap.gov/launch/). The same DEM was used to define the stream network. The parameters required for simulating streamflow, as well as performing sediment yield using the MUSLE soil erodibility K factor, were parameterized within each interface using the Soil Survey Geographic (SSURGO; http://websoilsurvey.nrcs.usda.gov). Streamflow simulation required soil chemical, physical, and hydraulic model inputs, including maximum rooting depth, soil hydrologic group, moist bulk density, soil profile depth, saturated hydraulic conductivity, available water capacity of the soil layer, and soil texture data (% clay, sand, silt, and rock fragment content) (Figure 2). Surveys and reports on the study area were used to obtain land use and land cover information as well as
general land management data, including tillage types and dates, planting, fertilization, and harvests for most fields [unpublished data, Heidelberg University; [28] [29] ]. Daily weather data (e.g. minimum and maximum temperature, and rainfall) were obtained from the PRISM Climate Group, Oregon State University [44].
The USGS monitors water quality at the outlet of Rock Creek, 0.8 km (0.5 mi) from the confluence with the Sandusky River (USGS station 04197170) as part of the Heidelberg Tributary Loading Program (HTLP). The station has been in operation since 1982 [45] and is described in detail in Nelson et al. [28]. Since the two interfaces calculated slightly different areas for the watershed, the observed values were adjusted according to each interface’s calculation of watershed area (7576.54 ha for ArcAPEX and 7560.85 ha for NTT).
The APEX model was constrained with soft data, including the assurance that simulated values were within 15% of the average annual evapotranspiration (ET) and tile drainage (QDR) values of 524 mm [46] and 283 mm [43], respectively. Soft data are information on processes within a budget that may not be directly measured, including those found in literature, such as annual evapotranspiration (ET), tile drainage, crop yields, or certain species of nutrients [47]. The annual average yield ± 35% for corn, winter wheat, and soybeans was used to constrain crop yield data, which were taken from the Ohio Agricultural Statistics 2015 Annual Bulletin and 2009 Ohio Agricultural Statistics reports [48] [49].
2.2.3. Model Setup
The APEX model version 0806 [37] [50] was used in this study. It is important to note that although the NTT interface states that it uses APEX 0806, the executable has been modified. However, the modifications are not documented. ArcAPEX and NTT interfaces were each used to build one project. The APEX project was built (delineated) into subareas along with the corresponding stream network using ArcAPEX [32]. The subarea, APEX’s smallest modeling unit, is a function of land use and soil type. An area upstream and contiguous to the outlet at which the flow measurements were made was delineated using the automatic subarea delineation feature on the DEM. The land use, soils, and slope definition tool was used to define the categories appropriately. Using management and land use data collected by study area personnel [51] to define the subareas for creating files resulted in delineation of 29 subareas (Figure 1).
Because NTT cannot currently delineate subareas, the shapefile from the ArcAPEX delineation was used to build an APEX project with the NTT interface [35] [36]. Data on land use and management practices were populated using the NTT interface subsequent to delineation and selection of soil and weather data inputs. Management operations included, but were not limited to, crop type, tillage method, planting date, fertilizer type and amount, irrigation type and amount, harvest date. Operations were the same as those entered in the ArcAPEX interface. Subareas were manually routed using the routing scheme adopted from ArcAPEX.
The model was run after creation of all subarea files, creating a default model folder for each interface. The model folder includes all necessary input, control, and executable files along with the output files. The parameterization process was then used to edit and update input and control files. The drainage code (IDR) was set to 900 mm [43] to assign tile drainage to subareas with predominantly crop coverage and slopes < 3% [52]. The Hargreaves [53] method was used to estimate ET in both interfaces.
2.3. Model Evaluation
2.3.1. Sensitivity Analysis
Important model parameters for calibration were identified by performing a global sensitivity analysis (GSA) [54] that used variance-based sensitivity analysis to quantify the contribution of change in model parameters to the change in model outputs. The GSA also provided a flexible water simulation platform for incorporating different sets of model parameters. A GSA was implemented using the APEXSENSUN software [27] which is designed for Monte Carlo-based uncertainty analysis [55]. Defaults assigned by the respective interfaces were not altered for parameters that were not being tested for sensitivity.
Forty-two parameters related to nutrients and streamflow (and defined in [37] ) were tested through 20,000 simulations (i.e. 20,000 parameter combinations) for sensitivity. The standardized regression coefficient (SRC) was used as a GSA metric for streamflow, total phosphorus (TP), and total nitrogen (TN) predictions in the APEX model. Parameters in which SRC > 0.05 were considered sensitive. The sensitivity of parameters with an SRC > 0.05 for streamflow, TN, and TP simulation was determined based on the percentage bias [PBIAS]; [56] and Nash-Sutcliffe efficiency [NSE]; [57] performance measures. Sensitive parameters based on either NSE or PBIAS were selected and used during model calibration and validation. The equations for all simulated components are described in detail in the APEX model theoretical documentation (30).
2.3.2. Calibration and Model Evaluation
Previous [58] [59] and current literature review found that most studies used only statistical performance measures to determine adequate calibration and validation. While Wang et al. [60] recommends that modelers obtain a correct water balance that includes all hydrologic components (e.g. surface flow, subsurface flow, percolation, evapotranspiration) and crop yields, with crop yields as the absolute minimum criteria level if no measured water quantity data are available, few of the ensuing peer reviewed papers follow this recommendation rigorously. According to Nelson et al. [29], it is important to utilize the soft data to obtain realistic simulations of various management practices, thus ensuring one gets the right answers for the right reasons [61]. In this study, model performance was assessed using the NSE and PBIAS statistical performance measures calculated with APEXSENSUN [27]. The criteria thresholds for NSE and PBIAS used in this study were the same as those used by Nelson et al. [29]. Moriasi et al. [62] considered a model to be calibrated for streamflow if the NSE ≥ 0.50 and PBIAS ≤ ±15%, and for N and P if the NSE ≥ 0.35 and PBIAS < ±30%. In addition, the model was constrained during calibration using soft data [47] value ranges for ET, QDR, and crop yields described earlier. Long-term crop yield ranges (soft data) used to bound the parameter values for corn, wheat, and soybean were 8.6 ton ha−1 ±35%, 4.0 ton ha−1 ±35%, and 2.9 ton ha−1 ±35%, respectively [48] [49]. The statistical performance measures and the soft data constraints together are referred to as performance criteria throughout the rest of this paper. Comparisons on model simulation performance were made for each individual criterion. According to Nelson et al. [29], models evaluated on a daily time step did not meet the selected criteria when simulating daily streamflow. This could be attributed to the precipitation and streamflow measurement cutoff at midnight for each day and the lag time between a precipitation event and a streamflow surge. In a study to determine the impact of length of the calibration period on model performance, Nelson et al. [28] found that the model performed best at an annual temporal scale when using long term (25 years) data to calibrate the model. This study was performed in the same study area with the same 25 years of measured data. Therefore, model performance was evaluated at an annual time step.