StreamFlow Variability of 21 Watersheds, Oregon

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 |