You are on page 1of 19

Atmospheric Research 77 (2005) 388 406 www.elsevier.


A PCSWMM/GIS-based water balance model for the Reesor Creek watershed

D. Smith a,*, J. Li b, D. Banting c
b a ENSCIMAN Program, Ryerson University, Toronoto, Ontario, Canada Department of Civil Engineering, Ryerson University, Toronoto, Ontario, Canada c School of Geography, Ryerson University, Toronoto, Ontario, Canada

Received 19 April 2004; received in revised form 9 December 2004; accepted 10 December 2004

Abstract This paper presents the results of a study of a watershed experiencing the pressures of landuse change resulting from urban development. The study was undertaken to facilitate an understanding of the water balance of the watershed by developing and implementing watershed procedures that are to be addressed in a watershed plan. There were three components to the research: firstly, observation of the effects of spatially distributed rainfall measurements and their effect on modelling were assessed. Secondly, the model was then calibrated by observing how differing techniques can discretize both the landscape (e.g. land-use and soil type) and incoming precipitation. Finally, a modelling methodology was developed to integrate a Geographic Information System and a hydrologic model (e.g. Storm Water Management Model) in a water balance analysis on a watershed basis. Results show that, under certain conditions, kriging spatially distributed rainfall values can help predict rainfall at ungauged (virtual) sites. Discretization of a watershed was found to affect the differences between measured and generated runoff volumes; however, this can be refined with calibration. It was seen that a strong correlation between measured and predicted rainfall values did not always guarantee a strong relationship between measured and generated runoff Recommendations include the use of a longer time series of rainfall, streamflow and predicted rainfall to observe temporal variations, and the need to assess the differences in modelled rainfall values generated by various surface

* Corresponding author. E-mail addresses: (D. Smith), (J. Li), (D. Banting). 0169-8095/$ - see front matter D 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.atmosres.2004.12.010

D. Smith et al. / Atmospheric Research 77 (2005) 388406


interpolation methods (e.g. Inverse Distance Weighting and other kriging options) currently available in GIS packages. D 2005 Elsevier B.V. All rights reserved.
Keywords: Kriging; GIS; SWMM; Water balance; Watershed; Precipitation; Model; Lumped; Clustered; Grid

1. Introduction Modelling is one of the key approaches to watershed management, and one of the most compelling concerns in watershed modelling is the spatial and temporal detail used (Kelly and Wool, 1995). With the convergence of software currently available for stormwater management and for spatial analysis, modelling approaches becomes an important concern. The United States Environmental Protection Agencies Storm Water Management Model (SWMM) has gained widespread use and is now available in a PC version. It was designed to accurately simulate real storm events on the basis of rainfall (hyetographs) and other meteorological inputs, as well as system attributes (e.g. watershed, conveyance, storage/treatment) to predict stormwater runoff in terms of quantity and quality (James and James, 2000; Huber and Dickinson, 1992). Geographic Information Systems (GIS) are recognized for their ability to model and enable analysis of spatially distributed phenomena. For instance, opportunities are available in a GIS to enable interpolation among a set of point observations, with the user able to adjust assumptions about the effect of distances or directions between sample points (Burrough and McDonnell, 1998; Johnston et al., 2001; Burke, 1995). Procedures such as kriging assign weights surrounding measurements to predict a value for an unmeasured location (Vieux, 2001; Johnston et al., 2001), allowing the user to model bsurfacesQ representing spatially continuous interpolated values. Stormwater management would seem to benefit from this capability in providing spatial detail for analyses of precipitation, storage and runoff. In this study, the rainfall-interpolation opportunities of kriging were the focus. Kriging is based upon two hypotheses: (1) the mean is constant throughout a region and (2) the variance of differences is independent of position, but depends on distance (Vieux, 2001; Johnston et al., 2001). If the first does not occur, then a trend exists and must be removed to maintain kriging in its simplest form. The second suggests that a variogram (variance over increasing distance between observations) for the data set can be used to estimate the variance for the region. Ordinary kriging assumes the following: Z s l es 1

where Z (s ) is the measured value at i th location, l is an unknown mean constant and e (s ) is the error with the l constant. One of the problems with ordinary kriging is determining if the assumption of a constant mean is representative. Johnston et al. (2001) suggest that there is good evidence to support the rejection of this assumption. Nevertheless, as a simple prediction tool, it is considered flexible and relatively accurate.


D. Smith et al. / Atmospheric Research 77 (2005) 388406

A spatially dependent model such as kriging is based on properties of the variogram or most commonly the semivariogram, which depicts the square root of the variance over increasing distances. The semivariogram therefore quantifies the assumption in this study, that precipitation distant from any point will be less similar than precipitation that is closer and that the strength of the statistical correlation is a direct function of distance (Burrough and McDonnell, 1998; Johnston et al., 2001; Environmental Science Research Institute (ESRI), 2002). Semivariance is defined as: ch 1=2N hd RZ xi Z xi h2 2

where c (h ) is the estimated semivariance for the distance h , N (h ) is the number of measured point pairs in the distance class h , Z (. . .) is a measured value in (. . .). Kerry and Hawick (1997) discussed rainfall prediction as an example of meteorological irregularities that make processes like kriging appropriate for decision makers who need spatially distributed calculations (e.g. agriculture: likely crop yields based on rainfall distribution). In their case, the rainfall used for the kriging dataset was inferred from satellite imagery. Their results concluded that rainfall kriging could be regarded as an effective interpolation method for areas where rainfall was ungauged. Similar to Kerry and Hawick (1997), Mizzell (1999) compared the kriging results of 62 rain gauges operated by the Columbia Airport National Weather Service and Doppler surveillance radar (WSR-88D) data provided by the National Weather Service for seven large rainfall events. Her results showed that radar could not (yet) provide the spatial distribution of surface rainfall that was needed for many operational and research applications. This was because of the great variability in the intensity and distribution of precipitation. She recommended that more radar-gauge comparisons should be conducted covering a larger number of storms and that future research should be conducted on these results which include analysis of the radar level II base reflectivity data to determine the origin of error sources. While rain gauges present excellent temporal detail of rainfall at a particular location, unless used in a network, they do not reflect the spatial variability of a typical storm event. In contrast, radar presents a detailed representation of the spatial and temporal attributes of rainfall over a large area. Nonetheless, estimated rainfall rates are derived indirectly using reflection measurements and are therefore subject to errors that need to be corrected by some form of ground-truthing (e.g. rain gauges) (Sinclair and Pegram, 2004). When combined, rainfall measurements could improve the performance of hydrological models (Sinclair and Pegram, 2004; Velasco-Forero et al., 2004a,b). For instance, Velasco-Forero et al. (2004a,b) observed the effect of calibration errors of radar rainfall data where radar data and rain gauge data were merged using Kriging with External Drift (KED) and an automatic methodology to compute valid bidimensional correlograms. Rainfall values from two different radar data sets (one that overestimated and other that underestimated) and rain gauge measurements were estimated independently using this non-parametric technique. Results suggested that the merging technique was effective by improving estimations of rainfall where radar calibration errors occur. Comparable to Velasco-Forero et al. (2004a,b), Sinclair and Pegram (2004) also joined rain gauge and radar data using a conditional ordinary kriging merging technique. By

D. Smith et al. / Atmospheric Research 77 (2005) 388406


combining a mean field based on the kriged rain gauge data, with the spatial detail from the radar data, a reduction in rainfall estimates by the radar system was observed. This is because the variance of the estimate is reduced in the vicinity of the gauges where they are able to provide good information on the true rainfall field. To date, results are being used to improve the South African Weather Services (SAWS) network of weather radars, as an input component of a flood forecasting system. The three objectives for this research also seek to address these concerns: ! To observe the effects of spatially distributed rainfall measurements and their effect on modelling. ! To calibrate the GIS-based water balance model by observing how differing techniques can discretize both the landscape (e.g. land-use and soil type) and incoming precipitation. ! To develop a modelling methodology that integrates GIS and hydrologic models in a water balance analysis on a watershed basis.

2. Case studythe Reesor Creek watershed, Ontario, Canada The Reesor Creek watershed is located in southern Ontario 20 km northeast of Toronto and has 4 subwatersheds that drain an area of approximately 35 km2 into Lake Ontario (Fig. 1A). It is a rural watershed underlain with predominately loam soils in a temperate climate and, as a result, the land-use is dominated by agriculture. Approximately, 10.4% (3.6 km2) of the watershed is urban. The Reesor watershed was selected because a rain gauge (in the Town of Stouffville) and an outlet stream gauge (station 02HC039) are located in the catchment and hydrographs depicted a very good relationship between each station (Table 1, Fig. 1B). Calibration of the SWMM-based runoff model for this study was done using the Stouffville rain gauge. 2.1. Research scope The standard practice of water balance analysis based on SWMM has been to use a spatially blumpedQ representation for calculations. This entails aggregation of properties to a watershed-wide scope and has been imposed on previous analysts because of computational limitations, or because data was not available to populate a more spatially distributed model database (Vieux, 2001, p. 1). This is no longer the case where modelling databases are becoming increasingly widespread, and more discrete due to the introduction of detailed digital databases, GIS-analysis software and enhanced micrcomputer-based computational capacity to handle them. By studying lumped, clustered and grid modelling, the study has been be able to identify the limitations that arose from the three approaches (Burke, 1995) (Fig. 2). The lumped model consists of all four subwatersheds (1 through 4) of the Reesor Creek watershed. Parameters were based on majority characteristics of the watershed. The lumped parameter model was applied to each spatially variable rain gauge. The clustered model is similar to the lumped model in terms of parameters; however, the subwatersheds


D. Smith et al. / Atmospheric Research 77 (2005) 388406

Fig. 1. The Reesor Creek watershed, Ontario, Canada: (A) subwatersheds and (B) land-use.

were modeled with adjusted values of drainage area, average width and percent impervious. Using a grid to overlay the Reesor Creek watershed, a discrete analysis was conducted on each grid cell (Fig. 3). Each cell represented a 1 km2 and in total, 54 cells were needed to overly the Reesor watershed. Each cell was also given a virtual rain gauge using the cells centroid to assign a value of rainfall. The cells were also measured for area of landuse, soil type and imperviousness and the results would determine the SWMM parameters. The lumped and subwatershed models were further subject to a simplified seasonal analysis where the precipitation time-series was divided into summer and fall. Summer was considered May 1 to August 31, 1999 and fall was September 1 to November 31, 1999. Other than time series adjustments all parameters were left the same. 2.2. Rainfall kriging and PCSWMM model This study used the following software: Computation Hydraulics Internationals PCSWMM GIS 2002 and ESRIs ArcGIS-ArcMap. Thirteen federally owned rain gauges were also used for interpolation. Table 1 depicts some of the characteristics of all 13 stations. All stations were correlated to the Stouffville gauge since it is the only gauge that is located in the Reesor Creek watershed.

Table 1 General attributes of precipitation and stream gauging stations Gauging station Units Time-step Per month (m3 and mm) May June July August September October November Total R2 UTM coordinate Distance to Easting Northing Stouffville (km) D. Smith et al. / Atmospheric Research 77 (2005) 388406 643,673 4,868,411 5.0

Stream gauge 02HC039a m3/s Rain gauges Oshawa Markham Bowmanville Buttonville Airport Stouffvilleb Cherrywood Bedford Kimberly Udora Burketon Janetville Pontypool Tyrone Average





332,722 664,073



5,227,123 na

mm/day mm/h mm/day 1/10 mm/h mm/day mm/day mm/h mm/h mm/day mm/day mm/day mm/day mm/day

daily hourly daily hourly daily daily hourly hourly daily daily daily daily daily

74.5 38.3 73.8 48 58.2 46.2 53.5 43.4 79.8 84 67.2 63.4 60.9

78.7 90.9 69.8 89.2 77 69 92.8 61.8 135 68.4 69.6 48 79.2

62.7 13.2 51.7 60.4 60.8 52.8 38.5 24.6 106 55.2 120 108 70 63.4

61 19 69 70 63 21 82 67 63 60 62 62 67 59

147 113 143 83.4 117 48.2 50.3 110 131 120 137 120 117 111

79.7 89.5 76.2 77 83.5 24.2 60 54 77.8 106 95.8 93.1 25.8 72.5

85.4 99.9 104 80.9 101 24.2 107 59 101 100 107 108 89.7

589 464 587 509 560 286 484 420 693 595 659 491 392 517

0.647 0.438 0.757 0.859 1 0.129 0.606 0.757 0.512 0.713 0.632 0.663 0.457

669,171 640,919 685,257 631,051 640,784 651,327 658,897 637,134 644,839 676,741 681,742 690,258 682,553

4,864,705 4,855,378 4,865,922 4,858,217 4,869,977 4,857,270 4,842,807 4,838,075 4,901,879 4,880,115 4,898,635 4,885,793 4,875,519

28.7 14.1 44.8 15.3 0 17 29.9 32.5 31.9 37.6 50.2 51.6 42.4

All rain gauges except Buttonville Airport have a correlation less that 0.76. For this study, a breasonably goodQ or bgoodQ relationship was considered R 2 z 0.6. A bad relationship was R 2 V 0.55. There are also data gaps in both the Pontypool and Tyrone gauging stations. a All volume totals are from raw data files and have not been subjected to baseflow separation. This study used the separated runoff for PCSWMM model comparisons. b All rain gauges were correlated to the Stouffville gauge.



D. Smith et al. / Atmospheric Research 77 (2005) 388406

Fig. 2. Spatial discretization of hydrologic models.

A point layer representing daily rainfall that fell in September 1999 was generated and the measurements of each gauge were entered accordingly. Using ArcMaps Geostatistical Analyst , ordinary kriging was run on each layer where both hypotheses

Fig. 3. Grid overlay of virtual rain gauges.

D. Smith et al. / Atmospheric Research 77 (2005) 388406


Fig. 4. A raster ordinary kriging coverage layer depicting the distribution of predicted rainfall about the Reesor Creek watershed for September 6th, 1999.

previously mentioned were satisfied (average = 2.4 and R 2 = 0.0332). The new kriged prediction layer was exported and labelled in context with the point layer (1 day/layer) being kriged (Figs. 4 and 5). A user defined 1 km2 grid was overlaid the Reesor watershed and the centroid of each grid cell was determined using a VB script termed get_centroid and applied using ArcMaps macro-option (Fig. 6A). The result was fifty-four virtual rain gauges, each gauge being characteristic of its applicable cell. The layer was exported and organized in to a time series,

Fig. 5. Rainfall kriging process used in GIS for each rainfall event.


D. Smith et al. / Atmospheric Research 77 (2005) 388406

Fig. 6. (A) Virtual rain gauge development and (B) kriged rainfall values were joined to each virtual gauge.

which was used in PCSWMM. The land-use, soil type, cell number and virtual rain gauge were joined and its attributes exported for use in PCSWMM. Also exported was the land-use and soil type attributes for the lumped, clustered and grid catchments. In the PCSWMM environment, lumped, clustered and grid models were generated for the Reesor Creek watershed using the data exported from ArcMap. SWMM parameters (e.g. infiltration rate, depression storage, etc.) were determined from literature (Huber and

Fig. 7. Runoff flows measured at station 02HC039 and generated by PCSWMM using the Stouffville rain gauge.

D. Smith et al. / Atmospheric Research 77 (2005) 388406


Dickinson, 1992; James and James, 2000) and the ArcMap exports. The purpose here is to apply a traditional calibration method throughout the development of the model, which in turn keeps the model consistent (Choi and Ball, 2002). Fig. 7 identifies flow (aggregated from hourly to daily, to match the time interval for precipitation) for the station 02HC039 and generated runoff by PCSWMM using Stouffville stream and rain gauge. The Stouffville gauge was chosen because it best represented precipitation falling on the Reesor watershed and it was located approximately 5000 m from the stream gauging station 02HC039.

3. Results and discussion Results from the September 1999 models can be seen in Table 2. PCSWMM has the capability to generate hourly runoff hydrographs using daily rainfall. Because of the
Table 2 Summary of lumped, clustered and grid models for September 1999 Station Oshawa Markham Bowmanville Buttonville Airport Stouffvillea Cherrywood Bedford Kimberly Udora Burketon Janetville Pontypool Tyrone Station Oshawa Markham Bowmanville Buttonville Airport Stouffvillea Cherrywood Bedford Kimberly Udora Burketon Janetville Pontypool Tyrone

Lumped vol. (m3) 334,375.6 254,909.1 337,648.5 206,156.2 267,773.4 41,960.1 85,860.2 226,022.9 286,383 245,206.5 286,757.7 244,798.8 242,170.8 Clustered vol. (m3) 272,174 245,586.8 308,761.9 230,645.2 248,266 33,968.5 69,266 180,486.8 229,183.1 196,344.2 228,919 196,263.2 193,923.3

02HC039 vol. (m3) 258,728.6

% Diff. 29.2 1.5 30.5 20.3 3.5 83.8 66.8 12.6 10.7 5.2 10.8 5.4 6.4 % Diff. 8.5 2.1 23.1 8.0 1.0 86.5 72.4 28.0 8.6 21.7 8.7 21.7 22.7

Lumped vol. (m3) 334,375.6 254,909.1 337,648.5 206,156.2 267,773.4 41,960.1 85,860.2 226,022.9 286,383 245,206.5 286,757.7 244,798.8 242,170.8 Clustered vol. (m3) 272,174 245,586.8 308,761.9 230,645.2 248,266 33,968.5 69,266 180,486.8 229,183.1 196,344.2 228,919 196,263.2 193,923.3

Grid vol. (m3) 237,220.5

% Diff. 41.0 7.5 42.3 13.1 12.9 82.3 63.8 4.7 20.7 3.4 20.9 3.2 2.1 % Diff. 14.7 3.5 30.2 2.8 4.7 85.7 70.8 23.9 3.4 17.2 3.5 17.3 18.3

02HC039 vol. (m3) 250,753.4

Grid vol. (m3) 237,220.5

Calibration was completed using the Stouffville station.


D. Smith et al. / Atmospheric Research 77 (2005) 388406

time-series differences between input rainfall and stream flow data, matching time to peak and event hydrographs was not a concern in this study. The focus was to generate a monthly water balance using only total volume in all three models. Attention was orientated towards developing a modelling protocol; however, further study should utilize matching input data time-series. In this case, all PCSWMM model outputs were in hourly intervals and volume balances for September 1999 was completed in order to compare the models. Observations concluded that the grid model when compared to the lumped, clustered and gauge 02HC039 increased the percentage difference. The discretization of the watershed and changes in the PCSWMM parameters of percent imperviousness, area and average width are what contributed to the differences. 3.1. Lumped model It was noticed that the Cherrywood station had the highest volume percentage difference, despite it being the closest rain gauge to Stouffville (17 km). One would assume that the sets of modelled results would be similar because of their location to one another; however, the difference may have been a result of measurement error, improper maintenance and natural/structural influence (e.g. elevated topography). Generally, most of the gauges correlated with the Stouffville gauge from average to well (0.6 to 0.9, respectively). Two of the stations (Pontypool and Tyrone) had gaps in the data set (Table 1, previous). In both cases, the Pontypool and Tyrone measured significantly less rainfall (greater than 100 mm) than the Stouffville gauge. The lower rainfall volume generated far less runoff in the PCSWMM environment. This is why both Pontypool and Tyrone generated smaller flows than the measured flow at 02HC039. In most cases (except Janetville and Pontypool), the percentage difference was greater than 15%. Several of these gauges had a reasonably good correlation with the rainfall measured at the Stouffville station (R 2 z 0.6). However, despite a moderate correlation between some gauges, it is the volume differences that contributed to most of the differences. For example, since PCSWMM is precipitation driven, the runoff volume generated by PCSWMM will reflect the volume of the measured rainfall. This response is further discussed later using Buttonville as an example. It was observed that there was no relationship between the spatial distribution of the rain gauge network and its correlation with the Stouffville gauge. 3.2. Clustered model The observations and discussion in the clustered model are very similar to the lumped model. However, the clustered model generally increased the percentage difference with the discretization of the Reesor Creek watershed. While the rainfall used in the clustered model was the same as the rainfall used in the lumped model, the difference here was the parameter adjustments used to represent subwatersheds 1, 2, 3 and 4. In this case area, average width and percentage impervious were calculated, and unlike the lumped model where each parameter represented the total area (e.g. percentage impervious). For example, the percent imperviousness for the lumped watershed was 8.41%, while the

D. Smith et al. / Atmospheric Research 77 (2005) 388406


clustered model was 10%, 15%, 2% and 2% for subwatersheds 1, 2, 3 and 4, respectively. Changes in imperviousness influence runoff results in PCSWMM since the SWMM engine is very sensitive to changes in this parameter (Fig. 8). The changes in catchment area and width are also sensitive parameters in the SWMM engine, which can increase or decrease the volume of runoff generated by PCSWMM. Nonetheless, for this study, the square root of the subwatershed was used to represent the average width, and the percentage difference may have been improved if changes in the average area during the calibration runs were conducted. 3.3. Seasonal model There is a drastic increase in percentage difference in the summer model. Because PCSWMM is best suited for urban watersheds, the rural landscape of Reesor Creek may be the reason for the differences. During the dry summer months the soil will dry out and increase soil moisture storage, therefore, during a storm event, rainfall is used up as soil recharge and retained as pervious depressions storage and not immediate runoff. On the other hand, the fall model generally underestimated flows. This may be because the fall is generally a wet time of year with frequent rainfall events. With the soil recharged, little infiltration will occur. The result can be overland flow that directly discharges to Reesor Creek as runoffthus, increasing runoff volumes. The condition of the watershed in the fall was similar to pavement or impermeable surfaces. Nonetheless, PCSWMM still underestimates the true measured volume because the percent impervious runoff is much less than the contributions of the entire watershed as groundwater flow and overland runoff.

Lumped and Clustered Percent Differences



Percent Difference





lle a m ra ille ille l oo yp Po nt Ty aw or oo or kh a Ki m be do vi to Ai rp St o Bu rk e m an O rry Be ar ne tv sh uf w df U ro fv ne t d d rly n Ja






Bu t

Station lumped clustered

Fig. 8. Percentage difference for both the lumped and clustered models.


D. Smith et al. / Atmospheric Research 77 (2005) 388406

3.4. Grid model The grid model discretized the Reesor Creek watershed into 1 km2 grid cells. Observations concluded that the grid model when compared to the lumped, clustered, and gauge 02HC039 increased the percentage difference. Similar to the discussion for the clustered model above, the discretization of the watershed and changes in the PCSWMM parameters of percent imperviousness, area and average width are what contributed to the differences. While this is not necessarily incorrect, since one would assume that the more discrete a model and its database is the more, so to should its percentage difference. Table 3 summarizes the results from the four models using flow generated by the Stouffville rain gauge. The table clearly depicts an increase in difference as the watershed was discretized. For instance, an increase occurred when both the generated models were compared (lumped against grid and clustered against grid), in this case 3.5 to 12.9 and 1.0 to 4.7, respectively. 3.5. Model evaluation A strong correlation in rainfall between gauges did not necessarily suggest a strong correlation in generated PCSWMM runoff volume (e.g. Lumped Buttonville: R 2 = 0.87, percent difference = 28.2). This occurred at several instances in the time series, where measured rainfall was significantly different. A larger rainfall would generate significantly larger runoff in the PCSWMM model and vice versa (Fig. 9).
Table 3 Results for the lumped, clustered, seasonal and grid models generated by the Stouffville rain gauge Generated volume (m3) May to November 1999 Lumped lumped summer lumped fall Clustered clustered summer clustered fall September 1999 Lumped Clustered grid (lumped) grid (clustered) 1,300,474.4 399,135.2 901,339.2 1,283,342.8 323,033.6 960,309.3 02HC039 volume (m3) 1,289,609.4 303,580.6 986,028.7 1,234,248.5 285,297.2 948,951.3 % Difference

0.8 31.5 8.6 4.0 13.2 1.2

267,773.4 248,266 237,220.5 237,220.5 Grid volume (m3)

258,728.6 250,753.4 258,728.6 250,753.4 Lumped/clustered volume (m3) 237,220.5 237,220.5

3.5 1.0 8.3 5.4 % Difference

Generated (September 1999) Lumped vs. grid Clustered vs. grid

267,773.4 248,266

12.9 4.7

D. Smith et al. / Atmospheric Research 77 (2005) 388406


Fig. 9. Volume differences in PCSWMM generated flow despite strong rainfall correlations. The red arrows identify peak differences.

This observation is significant because it identifies two things: (1) rainfall is spatially variable and (2) decision making models generated using rain gauges from outside the watershed boundaries or a single gauge for a large watersheds on their own may not truly represent runoff using SWMM as a prediction tool. The first point is further supported when the correlation values are plotted against the distance from the Stouffville gauge (Fig. 10). In this case, observations suggested that there is no relationship between the strength of the correlation with the calibration gauge (Stouffville) and the distance from it. This poor relationship may also be a direct result of the poor rainfall measurements, lack of maintenance on the rain gauge or lost data. It is recommended that future studies confirm rainfall data accuracy by correlating to other local gauges before this assumption can be made. The regression values for all of the virtual rain gauge correlated to the Stouffville gauge resulted in an average R 2 = 0.9694 and are a direct result of the kriging. This is because the values were weighted using the Stouffville gauge during the kriging process. The closeness of the virtual rain gauges to the Stouffville gauge caused the strong correlation. This observation is significant since it suggests that all of the rainfall that occurred in the Reesor Creek watershed boundaries is uniformly distributed and that rainfall values only btaper offQ gradually with distance. This is not always the case in reality. For instance, Fig. 11 depicts a hypothetical scenario where rainfall values bordering a watershed have a high correlation. However, a single measurement taken in the middle of the area is much lower. The result would be a predicted layer that is reflective of all the measurements and


D. Smith et al. / Atmospheric Research 77 (2005) 388406

Relationship Between Raingauge Location and Correlation with Stouffville Gauge
1 0.9 0.8 0.7

Correlation (R )

0.6 0.5 0.4 0.3 0.2 0.1 0 0 10 20 30 40 50 60 y = -0.0013x + 0.668 2 R = 0.0085

Distance (km)

Fig. 10. Spatial variability of rainfall measurements.

includes the middle values. This is inferred to the relationship between the Stouffville gauge and the Cherrywood gauge (poor correlation R 2 = 0.128). Conversely, if the middle value was not measured or if the topography had a high elevation, then the prediction would be significantly different. In this case, the predicted value would be reflective of the gauges bordering it. This observation would also contribute significant percentage differences in generated SWMM runoff. An example of this scenario was tested on several poorly correlated gauges in this study (R 2 V 0.55); the results can be seen in Table 4. Table 4 suggests that more rain gauges should be used in this type of modelling to avoid bexaggeratedQ predictions. For this test, the kriged rainfall was generated without the gauges Cherrywood, Markham, Janetville and Tyrone. The kriging was then re-run with the gauges removed and measurements from the predicted layer were taken at the point at which they originally were. The new values were then correlated to the Stouffville gauge and, in all cases, the predicted values were observed to have a better correlation. However, the new values were not reflective of the

Fig. 11. Measured against predicted rainfall.

D. Smith et al. / Atmospheric Research 77 (2005) 388406 Table 4 A comparison of measured and predicted rainfall correlations Station Cherrywood Markham Janetville Tyrone Measured with Stouffville (R 2) 0.129 0.438 0.632 0.457 Predicted with Stouffville (R 2) 0.852 0.888 0.847 0.811


Predicted with measured (R 2) 0.606 0.006 0.901 0.949

originally measured values. This suggest that there is a need for many more rain gauges to be used in this type of modelling to best represent the watershed and avoid bexaggeratedQ predictions.

4. Conclusions and recommendations Below is a list of the key conclusions made for this research. A short discussion follows to elaborate several of these points. n GIS was useful in generating the land-use, soil type and rainfall attributes for use in PCSWMM. n Kriging rainfall does accurately predict rainfall at ungauged (virtual) sites only under the following conditions: (1) all neighbouring stations must have a good correlation and (2) stations measuring lower values than neighbouring stations should be confirmed for errors in measurements or other systematic influence. n The methodology outlined in this study does contribute to understanding the effect of spatial distribution on rainfall. n Rainfall variance is independent of distance. Rainfall variability does occur with spatial location. n A strong correlation between measured rainfall values does not confirm a strong relationship with generated runoff. n Changes in percentage imperviousness of a watershed will affect modelled runoff. n Disaggregating a watershed does induce differences; however, calibration of some watershed modelling parameters can limit this (e.g. watershed width). n Seasonal variability does affect hydrologic modelling results. As a methodological protocol, this study is uniquely applicable to rainfall-runoff analysis using PCSWMM and ArcGIS software. Integration of data file exchange was a critical link for this study. In this case, attribute tables for soil, land-use and virtual rainfall were generated by ArcMap and analyzed to develop the parameters for the input files for PCSWMM. PCSWMM and ArcGIS were chosen because of their relative popularity with most professionals. The steps outlined in this study can be customized and applied by others to not only generate discretized runoff volume, but to improve upon the variability of spatially distributed rainfall measurements.


D. Smith et al. / Atmospheric Research 77 (2005) 388406

This study also observed the effects of disaggregating a watershed based on lumped, subwatershed, seasonal and grid models. Three of the four models did generate a reasonable water balance over the full time-series; however, a longer time-series would most likely generate a better representation of the watershed. For instance, results found in this study did not generate a good water balance for the summer months. This is because PCSWMM generates flow usually with short lag time, characteristic of an urban watershed. In urban environments, runoff has a shorter lag time and PCSWMM takes this into consideration when generating flow. However, since most of Reesor Creek is pervious, during the dry summer months, the soil will dry out and increase soil moisture storage. The result during storm events is a soil moisture recharge (first) and not runoff. On the other hand, the results that were generated during the fall months did lower the range of percentage difference, however, PCSWMM in all cases underestimated flow. This might be because fall is generally a wet time of year with frequent rainfall events. With the soil being recharged, infiltration rates are lowered. The result can be overland flow that directly discharges to Reesor Creek as runofffar with little or no depression storage in the pervious area (which is the majority). However, PCSWMM is better suited for a watershed in this condition because saturated soils are similar to pavement or impermeable surfaces. Nonetheless, PCSWMM still underestimated the true measured volume because the percentage impervious runoff is much less than the contributions of the entire watershed as groundwater flow and overland runoff. It would appear that the discretization of the watershed did introduce volumetric differences into the models. The increase in data and parameters from lumped to grid scale modelling most likely became more representative of reality and less like a single parameter model. When the generated flows using the Stouffville gauge (calibration gauge) are compared, the percentage difference form lumped to grid ranged from 8.3% to 0.8%, all of which are still good results in terms of a water balance (less than 10% difference). The third objective guided observations to reflect the effects of spatial distribution on rainfall. In all cases, flow generated by PCSWMM did differ when using different gauges outside of the watershed. Using the Stouffville gauge to calibrate the Reesor watershed parameters of all four models, the remaining 12 gauges were run in PCSWMM and, in each case, differences were introduced. Because PCSWMM is precipitation driven, if there is no rainfall, no flow is generated. The inverse will occur if it does rain. Using kriging to improve the rainfall values to generate virtual rain gauges improved the rainfall measurements within the watershed. The virtual rainfall and rain gauges are more representative of the precipitation falling on the Reesor watershed. The kriged rainfall approach, while slightly time consuming to generate, lowered the percentage difference between spatially distributed rainfalls. It should be noted that even if the model is left as a lumped parameter model, and only the virtual rainfall was generated, the difference results still improved. This study shows the effect of lumped, clustered, and gird modelling. It has also reviewed the influence that spatial distribution has on rainfall. Together, the key objective was completedto demonstrate a new modelling technique using GIS and hydrologic modelling in a water balance. The new technique was to use the GIS to

D. Smith et al. / Atmospheric Research 77 (2005) 388406


discretize both the rainfall and watershed attributes using kriging and a grid layer analysis. Results suggest that the introduction of GIS to the modelling did introduce differences; however, it was only because the model became more discrete and less like a lumped model. While this study is simply a protocol for further research, it does answer the question how do space and time affected modelling? The following outlines some of the key recommendations that would improve future GIS-based hydrological analysis and serve as a guide for further research. n Rainfall kriging should be conducted for a longer time-series (e.g. 6 months to 1 year) and it would be preferred to observe if the grid analysis becomes erroneous with time. n To confirm data accuracy prior to analysis (e.g. rain gauge maintenance schedules, interviews, etc.). n To demonstrate the effectiveness of differing surface interpolation methods (e.g. Inverse Distance Weighting and other kriging options) on rainfall. n To apply the kriging model to a minimum of 50 gauges to improve the values generated by the prediction layers. n To assess the importance of using only localized rainfall measurements for PCSWMM models. The percentage differences generated depict the significant differences in a models prediction.

Burke, S., 1995. Land surface parameterization: regionalized verses distributed approach to groundwater discharge. Physics, Chemistry and Earth 20 (34), 331 337. Burrough, Peter A., McDonnell, Rachel A., 1998. Principles of Geographical Information SystemsSpatial Information Systems and Geostatistics. Oxford University Press, New York. Choi, Kyung-sook, Ball, James E., 2002. Parameter Estimation for Urban Runoff Modelling. Urban Water, pp. 31 41. Environmental Systems Research Institute (ESRI), 2002. Geostatistical Analyst ArcMap Help File: Semivariograms and Covariance Functions. Environmental Systems Research Institute, Support: ArcMap. Huber, Wayne C., Dickinson, Robert E. (USEPA), 1992. Storm Water Management Model Version 4: Users Manual. Environmental Research Laboratory Office of Research and Development U.S. Environmental Protection Agency Athens, Georgia 30613. EPA/600/3-88/001a. James, William, James, Robert C., 2000. Water Systems Models (1)HydrologyA Guide to the Rain, Temperature, and Runoff Modules of the USEPA SWMM4. Computational Hydraulics International (CHI), Guelph, Ontario. Johnston, Kevin, Ver Hoel, Jay M., Krivoruchko, Konstantin, Lucas, Neil, 2001. Using ArcGIS Geostatistical Analyst. Environmental Systems Research Institute, USA. (ESRI). Kelly, Martin, Wool, Tim, 1995. Linked Watershed/Water-body Model. National Conference on Urban Runoff Management: Enhancing Urban Watershed Management at the Local, County, and State Levels. Conference: March 30 to April 2, 1993, The Westin Hotel Chicago, Illinois. Centre for Environmental Research Information. EPA/625/R-95/003. Kerry, K.E., Hawick, K.A., 1997. Spatial Interpolation on Distributed, High-Performance Computers. Technical Report DHPC-015. Department of Computer Science, University of Adelaide, Australia. Mizzell, Hope Poteat, 1999. Comparison of WSR-88D derived rainfall estimates with gauge data in Lexington County, South Carolina. Thesis. University of South Carolina, Department of Geography.


D. Smith et al. / Atmospheric Research 77 (2005) 388406

Sinclair, Scott, Pegram, Geoff, 2004. Combining radar and rain gauge rainfall estimates for flood forcasting in South Africa. Proceedings from the Sixth International Symposium on Hydrological Applications of Weather Radar. Melbourne, Australia; 24 February 2004. Velasco-Forero, C., Cassiraga, E., Sempere-Torres, D., Sanchez-Diezma, R., Gomez-Hernandez, J., 2004a. Merging radar and rain gauge data to estimate rainfall fields: an improved geostatistical approach using nonparametric spatial models. Proceedings from the Sixth International Symposium on Hydrological Applications of Weather Radar. Melbourne, Australia; 24 February 2004. Velasco-Forero, C., Sempere-Torres, D., Sanchez-Diezma, R., Cassiraga, E., Gomez-Hernandez, J., 2004b. A non-parametric methodology to merge rain gauges and radar by kriging: sensitivity to errors in radar measurements. Proceedings of the Third European Conference on Radar in Meteorology and Hydrology, Visby, Island of Gotland, Sweden, 610 September 2004. Vieux, Baxter E., 2001. Distributed Hydrologic Modelling Using GIS. Kluwer Academic Press, Boston, USA.