Supplement 3: Upper Upatoi SWAT Model

Supplement 3: Upper Upatoi SWAT Model

A SWAT model was built for the Upper Upatoi Watershed for this comparison and to support work with the Department of Defense (DoD) to assess ecosystem services on DoD base land. The SWAT model was built using typical data sources both for model parameterization and for calibration.

1.Data Sources

The Upper Upatoi SWAT model used typical US data sources that were supplemented by some DoD-specific data sources. The data sources included:

1. DEM: A DoD-supplied DEM with 10m resolution was used to specify the watershed topography.
2. LandUse: The NLCD 2006 land use map [Fry et al., 2011] was used to map land cover. The NLCD land classes were associated with similar SWAT land use classes, as listed in table S3.T1. Each NLCD land class was associated with a distinct SWAT land use class to keep the land surface response distinct. SWAT land classes including SWRN were edited to have parameters more representative of local conditions (see S3.T2).
3. Soil: The model used soil that combined components from the SSURGO soil dataset [Soil Survey Staff, n.d.] and the STATSGO2 soil dataset [Soil Survey Staff, n.d.]. SSURGO soil data was preferred, but at the time of model construction only STATSGO2 soil data was available for Harris and Talbot Counties (the other counties associated with the Ft. Benning Army Post, Muscogee, Chatahoochee, Marion, and Stewart in Georgia as well as Russell in Alabama all had SSURGO soil data available). Maps of the STATSGO2 and SSURGO soil data were combined in ArcGIS to create a single soil layer keyed to soil MUID. In addition, the USERSOIL table in SWAT was updated to include the SSURGO descriptions of the soils located Ft. Benning counties and a lookup table from soil MUID to soil name was built. Soil classification was performed based on soil MUID using the lookup table to link soil MUID to soil name in the USERSOIL database.

While the soil data is generally good, there is a major “impacted area” in the S-center of the Upper Upatoiwathershed where soil depth is listed as 25.4 mm. We expect that soil data in that area is effectively unmeasured as a result of the impacts to that area, and thus not reliable;specifically, we expect that the soil is actually deeper and more like the surrounding land.

1. Precipitation and temperature: Precipitation and temperature data came from NOAA GHCN daily gauge USW00093842, located in Columbus, GA. This gauge is technically located outside of the Upper Upatoi Watershed, but it is the nearest gauge with a time record over the complete period of interest for this and ongoing DoD studies.

USW00093842 was chosen over other nearby gauges after examining the correlation of gauge precipitation with measured discharge to assess the usefulness of each gauge for rainfall-runoff modeling. In addition, the USW00093842 precipitation was compared with radar-based precipitation estimates from the NexRAD Stage IVprecipitation dataset [NCAR Earth Observing Laboratory, 2011]for the region supplied by Katie Price [Price et al., 2013]. The average annual rainfall of the radar-based precipitation estimates was 94% of the gauge estimates, and while there was some spatial structure to it (radar predicted 10-15% less precipitation in the N and NE, ~5% less in the south, and similar amounts in the west), the differences were not severe. The Nexrad Stage IVdataset only covers 2006 to the present, so the USW00093842 rain gauge was used to provide a longer, consistent time series.

Other climate estimation was done using the SWAT weather generator with the default US First Orderweather generation table.

1. Discharge: The Upper Upatoi SWAT model was calibrated to discharge at USGS gauge 02341800, which provides discharge estimates from 1968-present near McBride Bridge on the Ft. Benning Army Post. Gauge 02341800 is listed as having good data quality for all times outside the estimated discharge.

In addition, USGS gauges 02341750 and 02341665 were used to examine the estimates of discharge and water yield in subbasins of the Upper Upatoi watershed but were not used during calibration. These two gauges started recording in 2008 so the period of record is much shorter; gauge 02341750 is listed as having good data quality and 02341665 is listed as having poor data quality.

2.Model setup and edits

The Upper Upatoi SWAT model was built using ArcSWAT2012.10_1.8, and with the SWAT2012revision 591 executable. The model wasset up using many default settings, although a few were changed; in addition, various tree growth parameters and management settings in the model were edited to allow for significant leaf area accumulation in the forests which did notoccur under default parameters.

During watershed delineation, a flow accumulation area of 1500 ha was used to define subwatershed size. In addition, watershed outlets were placed at the location of the three USGS gauges listed above, resulting in a total of 39 subwatersheds.

HRU definition was done as listed above using a 2% threshold for land use, a 5% threshold for soil, and a single slope class, which results in 1270 HRUs.

The default Penman-Monteith PET calculation was replaced by the Hargreaves PET calculation since only observed temperature (and not observed radiation, relative humidity, or wind speed) were supplied.

In addition to these settings, several parameters for tree growth were adapted to allow for better leaf accumulation and, hence, better ET estimates. During early model runs, transpiration seemed much lower than we expected. Our investigation showed that this occurred because the leaf-area index in forest areas was very rarely near the maximum (making actual ET near PET), which occurred largely because by default forest land uses included a kill and harvest management practice at the end of each year. Therefore, we made the following parameter changes to get approximately correct seasonality for all forest land uses (FRSD, FRSE, FRST)

1. Set IGRO = 1 to simulate starting with a fully grown forest
2. Set LAI_INIT = 3.0 for a leaf area index of 3 for FRSD, FRSE, FRST
3. Set BIO_INIT = 1000 to start with a forest mass of 1000 kg/ha for FRSD, FRSE, FRST
4. Set PHU_PLT = 2485 for FRSD and FRST, PHU_PLT = 3500 for FRSE to start with almost no heat units left to achieve growth for the first year.
5. Remove harvest/kill management operation from management settings for FRSD, FRSE, and FRST to allow for ongoing forest life and buildup of forest biomass over time.
6. Add constant fertilization (fertilizer #11-52-00) on a daily basis with 3 kg of fertilizer/ha each day. This prevents tree growth from being phosphorous limited, which is not something that we believe occurs in this site (at least not until trees are much more mature than the young growth we see here).
7. Change PLTPFRs of FRSD, FRSE, and FRST to PLTPFR(1) = 0.0020, PLTPFR(2) = 0.0019, and PLTPFR(3) = 0.0018 to adjust the trees’ phosphorous demand at different stages of the growth cycle.

The above settings ensure that trees reach full LAI in the summer, but still provide appropriate seasonal variation with significant drops in LAI and result ET during the winter. The addition of fertilizer and changes in phosphorous demand assure that the trees are not nutrient-limited in their growth. We note that the addition of fertilizer does mean that nutrient outputs from our model are likely to be highly biased; however, since this study does not investigate nutrient cycling in the watershed we believe this is acceptable.

These tree settings were tested by examining the time series of tree LAI for the SWAT runs. With these tree settings in place we observed that LAI reached the maximum tree LAI during late spring and dropped off in the fall. This process resulted in forest HRUs transpiring at near PET when not water-limited in the summer, as expected. Reductions in PET in the winter are also expected, although further investigation of the amount of reduction for the local loblolly and long-leaf pine series should be done.

3.Model calibration

We performed a manual calibration to match discharge peaks, the sustained baseflow, and the observed recession curve concavity of the hydrograph from 1/1/2004-12/31/2011 at USGS gauge 02341800, which we considered the bottom of the Upper Upatoi Catchment. Initially automatic calibration was attempted, and automatic sensitivity analysis was run. However, the automatic calibration approach did a poor job of fitting the long, relatively flat baseflow in this region because they have much less power in typical calibration parameters than discharge peaks. Hydrograph separation approaches or log transforms could have solved this challenge but were not attempted for this model. A manual approach was selected instead, focused on previously found sensitive parameters and others expected to strongly affect the timing and amount of discharge at the catchment outlet.

The parameters selected (see S3.T2) were varied individually to improve the model fit in R2, Nash-Sutcliffe efficiency (NSE), and percent bias (PBIAS), as well as the appearance of the hydrograph fit as determined by human observation. Once a parameter stopped improving model fit, another was selected that seemed to be most likely, given the equation structure of the model, to improve the model calibration with the least adjustment, and this was continued until a sufficiently good model was built. A list of the parameters adjusted and their final values is shown below in S3.T2. Absolute adjustments mean that the value was replaced by the new value, relative adjustments mean that the factor listed was multiplied by the original value, and additive adjustments mean that the factor listed was added to the original value. Validation was then run in a separate time period without calibration to assess the ability of the model to perform over multiple time periods.

Manual calibration of the hydrograph resulted in a high-quality model fit, as shown in figure S3.F1. Calibration was performed from 2008-2011, with 2 warm-up years for the model; the calibration period has R2= 0.70, NSE = 0.69, and PBIAS = 6.6% for a daily timestep and R2= 0.91, NSE = 0.90 for a monthly timestep. Validation was performed from 2004-2007, also with 2 warm-up years for the model, to avoid any benefit of a self-consistent model state by following the calibration period. The validation period has R2= 0.59, NSE = 0.58, and PBIAS = -11.0% for a daily timestep and R2 = 0.73, NSE = 0.71 for a monthly.These statistics suggest that the model quality is good to very good [Moriasi et al., 2007].

Table S3.T1: Standard SWAT databases were used to translate the NLCD 2006 land use dataset into SWAT land use classes. SWAT land uses were adapted to local conditions as discussed in S1.2

S3.T2: Calibration was performed manually on the parameters listed below; parameters to calibrate were determined on the basis of an automatic sensitivity analysis.

S3.T3: The calibration and validation statistics of the SWAT model suggest that it is good to very good [Moriasi et al., 2007]

S3.F1 The observed and modeled hydrographs show that the model fits the observations quite well in most cases. While the peaks do not perfectly fit the observed peaks in all cases, the general peak timing is quite good, and the hydrograph recession from peaks and the long flat baseflows match well.