Integrated environmental modeling is inspired by modern environmental problems and enabled by transdisciplinary science and computer capabilities that allow the environment to be considered in a holistic way (Laniak et al., 2013). In an agricultural context, synthesis and quantification of multidisciplinary knowledge via process-based modeling are essential to manage systems that can be adapted to continual change (Ahuja et al., 2007). The Soil and Water Assessment Tool (SWAT) (Arnold et al., 1998) is an example of such a process-based model. It has been developed over the past 30 years to evaluate the effects of alternative management decisions on water resources and nonpoint-source pollution in large river basins through the simulation of major processes including hydrology, soil temperature and properties, plant growth, nutrient and pesticides dynamics, bacteria and pathogens transport, and land management (Arnold et al., 2012; Douglas-Mankin et al., 2010). Furthermore, a weather generator is included in the model to fill gaps that may exist in meteorological records.

The SWAT model has been extensively tested around the world for a wide range of hydroclimatic conditions, water and land management practices, and timescales (Douglas-Mankin et al., 2010; Arnold et al., 2012; Gassman et al., 2014). The wide adoption of the SWAT model has been prompted by preprocessing and post-processing software tools such as a GIS interface and extensive user documentation (Arnold et al., 2012), as well as several linked databases for crops, soils, fertilizers, tillage, and pesticides (Santhi et al., 2005). Among these, soil properties are especially important as they are needed for the simulation of influential processes such as evapotranspiration, soil water balance, nutrient dynamics, and sediment transport (Neitsch et al., 2005). However, the existing built-in database is only valid for SWAT applications in the USA. Accordingly, studies outside the USA require the development of a soils dataset by preprocessing available soils data into a format readable by SWAT, a time-consuming process as not all data required by SWAT are readily available for countries outside of the USA.

Worldwide, SWAT has emerged as one of the most widely used water quality watershed- and river-basin-scale models for simulation of a broad range of hydrologic and/or environmental problems (Gassman et al., 2014). These applications in different regions are described in the extensive body of peer-reviewed SWAT literature (Arnold et al., 2012). Specifically in Canada, the SWAT model has been used for hydrological simulations in most provinces, including Prince Edward Island (Edwards et al., 2000), New Brunswick (Chambers et al., 2011; Yang et al., 2009), Nova Scotia (Ahmad et al., 2011), Ontario (Asadzadeh et al., 2015; Rahman et al., 2012), Quebec (Lévsque et al., 2008), Manitoba (Yang et al., 2014), Saskatchewan (Mekonnen et al., 2016), Alberta (Mapfumo et al., 2004; Watson and Putz, 2014; Faramarzi et al., 2015), and British Columbia (Zhu et al., 2012). However, preparation of Canadian soils information in a consistent and usable format for SWAT is time-consuming (Rahman et al., 2012), as information has to be collected from soil reports and cross-checked against GIS datasets, missing soil variables have to be calculated from other physical and hydraulic properties, and all parameters have to be attributed to specific soil grids or polygons.

Some of this preprocessing work can be alleviated by using publically available databases that contain most of the information required by SWAT. The Soil Landscapes of Canada (SLC) database published by Agriculture and Agri-Food Canada (Soil Landscapes of Canada Working Group, 2010) is an example, and has been used in SWAT applications in Ontario (Asadzadeh et al., 2015; Rahman et al., 2012), Saskatchewan (Mekonnen et al., 2016), Alberta (Faramarzi et al., 2015), and British Columbia (Zhu et al., 2012). The SLC contains a GIS dataset series that provides information about the country's agricultural soils at the provincial and national levels. It was compiled at a scale of 1 : 1 million, and the information is organized according to a uniform national set of soil and landscape criteria based on permanent natural attributes (Soil Landscapes of Canada Working Group, 2010). The SLC encompasses the southern portions of the provinces of Ontario and Quebec and a larger portion of the Prairies provinces of Manitoba, Saskatchewan, and Alberta as far north as to the boreal shield. Coverage in the maritime provinces of New Brunswick, Nova Scotia, and Prince Edward Island is nearly complete (Fig. 1).

Although there are more detailed soil datasets available at provincial levels (e.g., AGRASID dataset in Alberta), selection of SLC for integration with SWAT was based on the fact that (i) it covers most soils across the agricultural regions of Canada in a single dataset; (ii) it has been used in regional studies in Canada, as described above; and (iii) it is more easily applicable to large-scale national studies as broad-scale datasets require reduced resources to prepare and process data (Moriasi and Starks, 2010). Modeling studies comparing the performance of a single model (calibrated and uncalibrated) but using soil datasets with varying spatial resolution in the USA (i.e., the State Soil Geographic database (STATSGO) compiled at 1 : 250 000 scale, and the Soil Survey Geographic database (SSURGO) with scales ranging from 1 : 12 000 to 1 : 63 360) also revealed that using either dataset produced comparable results (Mednick, 2008).

Besides the American databases (i.e., STASTSGO and SSURGO), the authors are not aware of any other effort to produce a similar dataset from a national soils database for specific use with SWAT, such as the one presented here for Canada. Past efforts in preparing a large-scale soils dataset for modeling include the standardization of the FAO–UNESCO, but this dataset was not optimized for SWAT and is presented at a much coarser spatial resolution (i.e., 1 : 5 000 000; Batjes, 1997). The SOTER (Soil Terrain) database is another initiative to provide a global soils dataset, which was intended to have a global coverage at 1 : 1 million scale but was later degraded to 1 : 5 million scale due to the lack of means (Dobos et al., 2005). However, SOTER is not optimized for SWAT use and requires some variables to be calculated or estimated to this end (Bossa et al., 2012). Other databases at continental scale, such as the HYPRES in Europe, only cover soil hydrologic properties (Wösten et al., 1999). At national level, only a few countries besides the USA and Canada have a soil electronic database (e.g., Australia, Brazil, and China; Shi et al., 2004; Cooper et al., 2005), while these data are not available in most countries (Cooper et al., 2005). The reduced number of available datasets, coupled with the technicalities involved in translating these datasets into SWAT format and the required variables not reported in them, contribute to the lack of large-scale soil databases fully compatible with SWAT. These limitations emphasize the novelty and importance of the dataset presented in this paper. Besides presenting a soils database ready to use in SWAT simulations in Canada, the present work provides a framework to support similar initiatives in other regions using data from global soil databases.

Due to the importance of the SWAT model for integrated environmental modeling in Canada, and the prominence of the SLC database as a potential input dataset for this model at a national level, the objective of this work was to offer a country-level soils dataset in a format ready to be used in SWAT simulations. The dataset was derived to provide over 20 parameter values for different soil types that are varied for each soil layer. It was prepared in a format suitable for use in the ArcSWAT version of the model, which is attributed to a grid or polygon-based soil map. Such a laborious preprocessing exercise is widely, but inconsistently, adopted in SWAT simulations reported in the literature. Finally, deficiencies in the dataset are also presented and discussed.

The SLC database (http://sis.agr.gc.ca/cansis/nsdb/slc/v3.2/index.html, last access: 20 June 2017) is structured as a component-based GIS layer, whereby a single polygon may contain several soil records. This structure is similar to that of the State Soil Geographic (STATSGO) database in the USA (Srinivasan et al., 2010). Such structure creates a one-to-many relationship, whereby the multiple soil components of a polygon are not spatially defined. The actual soil information in the SLC database is stored in a number of tables linked together through intricate relationships (Soil Landscapes of Canada Working Group, 2010). Among these, four tables are relevant for developing a dataset for SWAT applications:

The Polygon Attribute Table (PAT) provides the linkage between geographic locations (polygons in the SLC GIS coverage) and soil landscape attributes in the associated database tables (e.g., unique soil ID in the Soil Name Table (SNT) and respective number of layers in the Soil Layer Table (SLT)).

The SWAT soils information is stored in the “usersoil” table, located within the SWAT 2012 database in Microsoft Access format (i.e., SWAT2012.mdb). Each soil record is stored as a new record (i.e., row) in the table. Specific soil variables (Table 1) comprise the 152 columns of the usersoil table. The first column is an OBJECTID field assigning a unique identifier for each record. Columns two through six pertain to soil classification. The second column is the map unit identifier (MUID), which is used for mapping a collection of areas grouped by the same soil characteristics. A single MUID may describe different soil types, which are stored with a record counter in the third column (SEQN), while a soil identifying name (SNAM), a soil interpretation record (S5ID), and the percent of each soil component (CMPPCT) are recorded in the fourth, fifth, and sixth columns, respectively (Sheshukov et al., 2009). Columns 7 through 12 describe major soil properties pertaining to the soil record, namely, the number of layers (NLAYERS), the hydrological soil group to which that soil belongs (HYDGRP), the maximum rooting depth of the soil profile (SOL_ZMX), the fraction of soil porosity from which anions are excluded (ANION_EXCL), the potential of maximum crack volume of the soil profile expressed as a fraction of the total soil volume (SOL_CRK), and the texture of the soil layer (TEXTURE).

The next 120 columns starting from column 13 (i.e., columns 13 to 132) describe the information for each layer of the soil record. These columns are arranged in sets of 12 variables each for 10 possible soil layers. The variable NLAYERS indicates how many of these sets should be populated. Variables for any sets beyond NLAYERS should be assigned a value of zero. The variables included in each set of soil layers are the depth from the soil surface to the bottom of the layer (SOL_Z), moist bulk density (SOL_BD), available water capacity of the soil layer (SOL_AWC), saturated hydraulic conductivity (SOL_K), organic carbon (SOL_CBN), clay (CLAY), silt (SILT), sand (SAND), and rock fragment (ROCK) contents, moist soil albedo (SOL_ALB), erodibility factor (USLE_K), and electrical conductivity (SOL_EC). Beyond the columns describing layered soil information, there are 20 columns (i.e., columns 133 to 152) describing two variables (i.e., soil CaCO3 (SOL_CA) and soil pH (SOL_PH)) for 10 soil layers. These variables are not currently active in SWAT and are assigned a value of zero.

Despite its usefulness as a source of soil information for hydrological simulations, the SLC dataset is not assembled in a format readable by SWAT or other similar models. For example, SWAT stores all the properties for a specific soil record in a single row in the usersoil table, while this information is stored in the SLC as multiple rows in two different tables (i.e., SNT and SLT). Thus, the information contained in the SLT database has to be processed to satisfy SWAT's format requirements. In addition, all properties in the usersoil table are spatially defined, while those of SLC are often stored in a multi-polygon structure with no unique spatial identification. Variables required by SWAT and contained in the dataset presented here were either extracted from SNT and SLT, or calculated from the information therein. Some other variables were estimated from published values. Extraction or calculation of variables was done through an R code that imported both SNT and SLT, screened the data for missing records and missing SWAT-required information (data screening is described in Sect. 5), and sequentially populated unique soil records in the database. The next section describes how these variables were defined.

The variables in SWAT's usersoil table refer to record indexing and soil classification, as well as soil properties pertaining to the entire profile or specific layers. The variables in each of these groups are described in the following subsections. The usersoil table starts with a number of columns that define the database and soil classification variables, followed by soil profile and layer information, and inactive soil properties (Table 2).

Soil properties are inherently uncertain due to spatial variability and precision of measurement methods (Lacasse and Nadim, 1996). This uncertainty has direct implications for hydrologic simulations and their interpretation (Beven, 2011). The SWAT model simulations, therefore, are subject to the uncertainty of the soil properties used as input. For example, hydraulic conductivity is highly spatially and temporally variable (Hillel, 1998), and these uncertainties are very difficult to be avoided. The processing applied to the original SLC database in the present analysis did not introduce any further uncertainty to the variables reported by SLC (e.g., saturated hydraulic conductivity). There is, however, some uncertainty relating to estimated and calculated parameters. These uncertainties are discussed in this section, although their quantification is beyond the scope of the present work.

An example of introduced uncertainty is the moist soil albedo in the present dataset (0.10), which is the average of a range reported in the literature (Sect. 6.3). However, any value selected would have some uncertainty associated with it from a modeling standpoint because a single value cannot represent the variability in moist soil albedo as the soil dries up. This is a recognized problem when trying to represent spatially or temporally variable parameters (e.g., soil moisture) using a point measurement or single value in hydrological models (Beven, 2011).

Another example of uncertainty is the HSG calculations, which required a number of assumptions. For example, the shallowest annual depth to water table was unavailable in the SLC and therefore estimated based on the drainage class of each soil record. Also, the assumption of artificial drainage resulted in assignment of dual-class HSGs to the less restrictive one. An assessment of HSG in the USA indicated a standard error of about one HSG (Stewart et al., 2012), suggesting that classifying soils in the neighboring groups is not uncommon and that there is some uncertainty associated with those estimates.

The estimation of erodibility factor (K; Eq. 1) would also be subject to uncertainty. This is illustrated by the range of erodibility factors reported for a single soil textural class (Table 4). This variability can arise for different reasons. One already mentioned is the precision of the method used to determine the textural classes. A second one is the procedure used to calculate K. For example, Neitsch et al. (2005) present an equation that requires a soil structure code used in soil classification with many types and subtypes. Since this soil structure code is note reported in the SLC, an alternative relationship (Eq. 4; Sharpley and Williams, 1990) that does not require the soil structure code was used. This relationship was selected to avoid the added uncertainty from estimating the soil structure code.

Finally, one last variable worthwhile discussing in term of uncertainty is the available water capacity of the soil layers. This variable was estimated as the difference between field capacity and permanent wilting point. The procedure used here to estimate available water content (i.e., the difference between field capacity and permanent wilting point) follows the same procedure used by SWAT (Neitsch et al., 2005) and is described elsewhere in the soil physics literature (Hillel, 1998). Therefore, it would not introduce any further uncertainty. However, using the soil moisture content at -10 kPa to replace records with missing soil moisture content at -33 kPa (Sect. 6.3) would introduce some uncertainty in available water capacity for the replaced records.

Overall, prediction of uncertainty in regional hydrologic modeling and a careful input data discrimination analysis prior to calibration is unavoidable (Faramarzi et al., 2015). Especially in large transboundary river basins where a consistent soil dataset is not available from a single source, a careful uncertainty assessment provides information on data and model quality. Although the authors are unaware of SWAT hydrologic simulations in binational watersheds that use soil datasets from both the USA and Canada, maybe due to lack of large-scale datasets for Canada, it is expected that the model output is subject to the quality and quantity of both datasets. Some aspects contributing to this uncertainty are (i) possible discontinuity in the mapping units (i.e., polygons) between the GIS layers of both datasets, (ii) the soil record being mapped in multicomponent polygons in the GIS coverage (Soil Survey Staff, 1999; Agriculture and Agri-Food Canada, 1998), (iii) differences in soil taxonomy between the USA system (Soil Survey Staff, 1999) and the Canadian system (Agriculture and Agri-Food Canada, 1998) of soil classification, (iv) the methods used to measure/estimate the physicochemical variables, which may differ in accuracy and precision, and (v) the natural variability in the calculation of some variables that cannot be measured (e.g., HSG; Stewart et al., 2012). Given the number of aspects influencing trans-boundary uncertainty and the large spatial scale of both the USA and the dataset discussed here, an assessment of such uncertainty is beyond the scope of the present study. However, this assessment is suggested to quantify the share of errors from these sources in hydrologic model projection in both upstream and downstream tributaries. These are the subjects of our continuing studies.

Although the SWAT database is in a proprietary format (i.e., Microsoft Access), the present soils dataset has been published in a nonproprietary format (i.e., comma-separated values (CSV) file) that can be opened in a variety of software packages. However, the dataset can be easily imported into the SWAT soils database using an automated import routine in Microsoft Access (Fig. 4). This import process consists of opening the SWAT2012 database and using the “Import Text File” tool under the “Import & Link” section of the “External Data” tab to read the CSV file. This action will prompt a window where the user can select the path to where the present dataset is stored and specify how and where the data are stored in the database. The option “Append a copy of the record to the table” should be selected, which activates a drop-down menu from which the usersoil table should be highlighted. Once these options have been processed, an “Import Text Wizard” window will be prompted, where the option “Delimited – Characters such as comma or tab separate each field” should be selected. Processing of this selection will prompt another window where the option “comma” should be automatically selected by the wizard. However, the user should activate the box “First Row Contains Field Names” since the first row of the present dataset contains the variable labels. Confirming the processing of the next windows should finalize the import process, and the data should be ready to be used in SWAT predictions.