Streamflow Variability of 21 Watershed Basins in Oregon Donnych Diaz and Tracy Ryan Portland State University March 17, 2010
Streamflow Variability of 21 Watershed Basins in Oregon
Donnych Diaz and Tracy Ryan
Portland State University
March 17, 2010
Abstract
Streamflow runoff was studied within twenty-one Oregon watersheds. The streamflow runoff data used in this study consists of monthly mean runoff values for each watershed ranging back as far as 1958. Slope, aspect, elevation, and land cover data were analyzed to determine which physical aspects of these watersheds affected streamflow runoff. Regression models were run using SPSS software and analyzed to determine both if the model meets the assumptions of ordinary least squares regression (OLS) as well as if the model was statistically significant. As part of the regression, the data was grouped into summer and winter month data, and then transformed by square root to meet the assumptions of OLS. The resulting analysis indicates that the model is more effective during the winter months when precipitation is higher. As compared to a summer R-squared value of .015 and an F-test with a significance value of .4, the winter R-squared value is more significant at .673, and the F-test is highly significant at .000. For the regression on the winter data, all independent variables other than slope are significant. Although it is suggested by the winter regression that elevation, land cover and aspect do have a correlation to streamflow runoff, more analysis is necessary to determine if this is an accurate assessment. The low statistical significance of the summer regression in particular indicate that other variables, such as soil cover and precipitation, affect streamflow runoff and should be considered in a predictive model.
2 |
INTRODUCTION
Of the many effects of global warming that have been analyzed by scientists, change in
streamflow runoff within watersheds is one that is not currently well understood. Most current
scientific thinking suggests that as the climate warms, decreased amount of snowpack lowers
the runoff rate of streams and rivers (Luce and Holden, 2009). As river systems can be
particularly complicated, this leaves questions as to what other variables have an effect on
runoff. This study aims to explore those questions, and looks specifically at the physical traits of
the studied watersheds and how they relate to streamflow runoff.
Twenty-one Oregon watersheds were chosen and analyzed for this project. The goal
was to take land attribute variables such as elevation, slope, aspect and land cover and create a
model using multiple regression where these factors could be used to predict resulting
streamflow runoff within these watersheds. Variables such as snowpack and precipitation were
left out to focus the model on physical, more slowly changing factors only. For this study, the
null hypothesis is that the physical attributes do not have an effect on streamflow runoff, while
the alternate hypothesis is that that the physical attributes do have an effect on streamflow
runoff.
The first step in this project was to collect data. A significant portion of the data used
was compiled from the United States Geologic Survey (USGS) as a digital elevation model
(USGS, 2009). From this data, information about slope, elevation and aspect was calculated
using GIS. Streamflow runoff data and land cover data were also obtained from the USGS. The
streamflow runoff data was aggregated as monthly means for each area, and go back to either
1958 or 1975 depending on the watershed. Land cover data was converted into a Land Cover
Roughness Factor (LCRF) using Manning’s Roughness Coefficients to assess how different types
of groundcover allow water to flow more or less efficiently within these watersheds. Once
these data were analyzed and aggregated, they were compiled into a database and shapefile
and SPSS software was used to create a multiple regression model.
3 |
STUDY AREA
Map 1: The 21 Oregon watersheds studied. The Year column below the map indicates the year to which data
was obtained (USGS, 2009).
The study area encompasses the 21 watersheds shown in Map 1. They are located
primarily along the North-South Willamette Valley corridor in Oregon and 3 are in the Eastern
part of the state. Selection of these watersheds was based on available streamflow data from
USGS National Water Information System that consisted of 52 years of mean runoff
measurements for more than half of the watersheds and 35 years for the remaining. The
majority 18 of the 21 watersheds are located on the windward side of the Cascade mountain
range. Precipitation levels west of the Cascades are between 1 to 5m annually, whereas east of
the Cascades, levels only reach between .250 to .500m annually (Broad and Collins 1996). The
4 |
precipitation level therefore is greater in significance in the majority of the watersheds than in
the 3 outliers.
The physical characteristics varied with a minimum to maximum elevation range
between 15m to 3,395m and a mean elevation range of 541m to 3,171.5m. There is a
predominately south to southwest aspect and slope range from 103% to 442% in percent (rise
over run) (Map 2)(USGS, 2009); mean slope ranged from approximately 13% to 184% (Map 3)
(USGS, 2009). The topography encompasses the following land cover types: barren land,
cultivated crops, deciduous forest, opens space developed, low to high intensity developed,
emergent herbaceous wetlands, evergreen forest, hay/pasture, herbaceous, mixed forest, open
water, perennial snow/ice, shrub/scrub and woody wetlands (Map 4) (USGS, 2009).
Map 2: Example of mapped aspect data, grouped watersheds 6 through 10. Of note is the prevalence of the red and brown colors indicating the prevalence within the watersheds is south by southwest (USGS, 2009).
5 |
Map 3: Example of mapped slope data, image is grouped watersheds 6 through 10. The slope data presented here is in degrees for easier visualization. Calculations were done with slope in percent form (USGS, 2009).
Map 4: Example of mapped land cover data. Image shown is the cluster of watersheds 6 through 10 (USGS, 2009).
DATA AND METHODS
6 |
The streamflow runoff data from the USGS National Water Information system was the
base dataset to which physical attribute data was added as the independent factors to be
analyzed against mean streamflow runoff. The physical attributes consisted of land cover,
elevation, slope and aspect. The 2001 NLCD (national land cover dataset) shapefile was
acquired from the Multi-Resolution Land Characteristics Consortium. Using GIS, the elevation,
slope and aspect were derived from DEMs from USGS Seamless Server, 1 arc second, 30m
resolution. Mean runoff data was collected for the periods of January through December from
1958 to 2008 for each watershed then averaged for the number of years. The coefficient of
variance, mean winter and summer flows as well as mean summer flow over annual flow were
calculated. In order to assess if there is a linear relationship between streamflow runoff and
the independent physical variables a statistical multivariate regression analysis was completed
using SPSS. This allows for the testing of a model to determine if any correlation exists between
the variables.
The multiple regression equation is shown in equation 1, where x is the independent,
explanatory variable; p is the number observations of the independent variables, and y is the
predicted value of the dependent variable.
y=a+b1 x1+b2 x2+…+bp xp (1)
A good model predictor, therefore, minimizes the sum of the squared residuals.
Using GIS, we derived zonal statistics for elevation, slope and aspect, which consists of a
mean, standard deviation, minimum, maximum, range and a total area for each watershed. The
land cover dataset was also derived using zonal statistics resulting in a total area per land cover
type per watershed. A Land Cover Roughness Factor (LCRF) was calculated in order to weigh
the effects of varying land cover types. This LCRF was derived by using the roughness ratio
portion of Manning’s Velocity formula to compute overland velocity shown in equation 2.
(Asante et al., 2007)
Velocity = 1/ManningN * RH 2/3 * √ Hillslope (2)
(ManningN is the Manning Roughness coefficient for the land cover and 1/ManningN is the roughness ratio)
7 |
The roughness coefficients are those used in Geospatial Stream flow models (GeoSFM) (Asante
and others, 2007) based on the land cover type as shown in table 1.
Table 1: Manning’s roughness values used for various land cover classes in GeoSFM
Anderson Code Description Manning Roughness
100 Urban and Built-Up Land 0.03
211 Dryland Cropland and Pasture 0.03
212 Irrigated Cropland and Pasture 0.035
213 Mixed Dryland/Irrigated
Cropland and Pasture
0.033
280 Cropland/Grassland Mosaic 0.035
290 Cropland/Woodland Mosaic 0.04
311 Grassland 0.05
321 Shrubland 0.05
330 Mixed Shrubland/Grassland 0.05
332 Savanna 0.06
411 Deciduous Broadleaf Forest 0.1
412 Deciduous Needleleaf Forest 0.1
421 Evergreen Broadleaf Forest 0.12
422 Evergreen Needleleaf Forest 0.12
430 Mixed Forest 0.1
500 Water Bodies 0.035
620 Herbaceous Wetland 0.05
610 Wooded Wetland 0.05
770 Barren or Sparsely Vegetated 0.03
>800 Tundra, Snow or Ice 0.05
The area per each land cover type for each watershed was then multiplied by the roughness
ratio and summed to obtain the total LCRF; where the larger the 1/ManningN ratio, the greater
the overland velocity and thus the greater the LCRF.
8 |
Initially the mean data for the dependent and independent variables were used without
transformation. The first regression using the January mean streamflow data showed extreme
non-linearity on the scatter plots between the variables suggesting the need for a
transformation. The variables were transformed using Log and Log10 with the same non-linear
results. The dependent and independent variables were then transformed by square root,
resulting in the best linearity between the variables. The regression model created uses the
summer (Jun.-Sep.) and winter (Dec. -Feb.) mean streamflow data and was tested for the four
assumptions of multivariate regression: linearity, constant variance, normality and
multicollinearity.
RESULTS
The resulting regression models had contrasting results for the summer and winter
dependent variables. Below in Tables 2 through 4 are the model summary statistics for both
dependent variables.
Table 2: Model Summary
Dependent Variable R R Squared Adjusted R
Square
Std. Error of
the Estimate
Durbin-
Watson
Summer .461 .212 .015 4.59257 2.026
Winter .859 .738 .673 4.99749 1.969
The winter model was a better predictor of streamflow runoff than the summer model with an
R square of .738 and adjusted R square of .673. The ANOVA test showed similar results:
Table 3 : Summer ANOVA
Model – Summer Sum of Squares Df Mean Square F Sig.
1 Regression 90.882 4 22.721 1.077 .400
Residual 337.467 16 21.092
9 |
Total
Table 4: Winter ANOVA
Model – Winter Sum of Squares Df Mean Square F Sig.
1 Regression 1126.636 4 281.659 11.278 .000
Residual 399.598 16 24.975
Total
The analysis of variance resulted in an F test value of 11.278 at a p=.01 significance level for the
winter model indicating that the null hypothesis can be rejected for the winter variable. The
summer variable was not significant with an F test value of 1.077 at a p = .400, therefore the
null hypothesis cannot be rejected. The coefficients of the variables for the summer model
indicated significance for only the land cover roughness factor with a t test of -1.926 and a
significance at .072 (Table 3). In the winter model the coefficients of the variables indicated
significance for the aspect, elevation and land cover roughness factor at a p = .01 (Table 4).
Table 5 : Summer Coefficients
Model –
Summer Unstandardized Coefficients
Standardized
Coefficients
t Sig.
Collinearity
Statistics
B
Std.
Error Beta Tolerance VIF
1 (Constant) 33.782 28.609 1.181 .255
Srslp=Slope -.509 .563 -.211 -.903 .380 .898 1.113
Srasp=Aspec
t
.080 1.635 .011 .049 .962 .961 1.041
Srelev=Elev. .312 .285 .301 1.096 .290 .654 1.530
Srlcf=LCRF -9.005 4.677 -.538 -1.926 .072 .630 1.586
10 |
Test for normality was done using the histogram for both the summer and winter
variables as shown in Charts 1 and 2. The distribution for both is trending towards
normal
Chart 1: Summer distribution Chart 2: Winter distribution
11 |
Table 6 : Winter Coefficients
Model – Winter Unstandardized
Coefficients
Standardized
Coefficients
t Sig.
Collinearity Statistics
B
Std.
Error Beta Tolerance VIF
1 (Constant) 43.820 31.13
1
1.408 .178
Srslp=Slope .183 .613 .040 .299 .769 .898 1.113
Srasp=Aspect 3.287 1.779 .241 1.848 .083 .961 1.041
Srelev=Elev. -1.156 .310 -.590 -3.728 .002 .654 1.530
Srlcf=LCRF -8.932 5.089 -.283 -1.755 .098 .630 1.586
Testing for constant variance, we generated a scatter plot of the studentized residuals and the
unstandardized predicted values for both the summer and winter regressions (Chart 3, Chart 4).
There is no pattern to the independent variables in either model, indicating both meet the
assumption of constant variance.
Chart 3: Summer Constant Variance Chart 4: Winter Constant Variance
The Durbin Watson values are 2.026 for the winter model and 1.969 for the summer model.
The summer model shows a very slight autocorrelation since its value is above 2.
Testing for linearity was done using the scatter plot of the dependent variables and the
independent variables. The winter mean variable indicates a better linear correlation among
the independent variables than the summer mean variable (Chart 5, Chart 6).
Chart 5: Summer Linearity Chart 6: Winter Linearity
12 |
DISCUSSION
DISCUSSION
The gauging and measurement of hydrologic processes is complicated at best. Given
the recent accumulated data evidencing the decrease in snowpack and earlier spring runoff
(Luce and Holden, 2009), accounting for the numerous factors contributing to these changes
can be daunting. We looked at a few of those contributing factors to determine if there is a
correlation between mean streamflow runoff and the physical attributes for 21 watersheds.
The temporal span of the data facilitated the calculation of mean streamflow from 35-52 years;
providing the analysis with population statistics. Although a lot of literature focuses on
hydrologic processes that are continuous data, the discrete data we examined, we argue could
have a direct correlation to the changes occurring in mean streamflow runoff.
The model predictors for the summer dependent variable proved inconclusive. We
attribute the poor result of this model to the decrease in precipitation during this time period
(Jun. – Sept ). Correlation between mean summer streamflow and the independent variables
are insignificant except for the LCRF; inferring that LCRF during periods of minimal precipitation
has a greater significance than slope, aspect and elevation. The performance of the winter
model is statistically and significantly better than the summer model. Again, we have to qualify
this by inferring that the model’s performance is due to the increase in winter precipitation.
Three out of the four independent variables are at significant levels except for slope during the
winter season. This may seem to be counter intuitive, but given the precipitation levels during
this time period, slope, per our model predictor, has little affect on mean streamflow runoff.
The winter model therefore can be a good predictor, however, like all regression analysis, other
contributing factors are not being taken into consideration.
13 |
CONCLUSIONS
Based on the significance statistics of our summer and winter regression models, this
regression model is not accurate enough to be a totally reliable predictor of streamflow in these
watersheds. However, important information can be gleaned from these models, and overall,
the models do suggest that certain of the studied variables do play an important role in
streamflow runoff variation. Land cover, being the only variable significant to a 90% confidence
level in both the summer and winter regression, should be explored more thoroughly as an
independent variable affecting streamflow. Elevation, with its strong beta value and high level
of significance in the winter regression, should also be explored further. The slope coefficient is
insignificant in both regressions, and this feature may be looked at for possible removal from
the model.
The lack of significance in the summer model indicates that more work should be done if
the model is to be predictive. In particular, the difference in significance between the summer
and winter regression models suggests that the water being put into the system in the form of
precipitation may be highly significant and should not be left out if the model is to be
predictive. As suggested in the literature, snowpack, as a factor influencing water input into the
system, should also be taken into account. Because the bedrock and soil geology in these
watersheds may have an effect on water absorption into the ground, these are also factors that
could be explored for possible correlations to streamflow runoff. Currently, the soil data for
Oregon is incomplete. Data was downloaded and analyzed, but did not spatially cover the
studied watersheds.
14 |
REFERENCES
Asante, K.O., Artan, G.A., Pervez, S., Bandaragoda, C. and Verdin, J.P. (2008) Technical Manual for the Geospatial Stream Flow Model (GeoSFM): U.S. Geological Survey
Open- File Report 2007–1441, 65 p.
Broad TM, Collins CA (1996) Estimated water use and general hydrologic conditions for Oregon, 1985 and 1990. USGS Water Resources Investigations Report 96 4080, Portland, Oregon
Fu, G., M.E. Barber, and S. Chen (2009) Hydro-climactic variability and trends in Washington State for the last 50 years, Hydrological Process, doi: 10.1002/hyp.7527.
Luce, C. H., and Z. A. Holden (2009) Declining annual streamflow distributions in the Pacific Northwest United States, 1948–2006, Geophys. Res. Lett., 36, L16401, doi:10.1029/2009GL039407.
United States Geologic Survey (USGS) (2009) http://seamless.usgs.gov/
15 |