GIS for Water Resources and Watershed Management - Chapter 2 pps

The performance of program DEDNM is evaluated by comparing the generated watershed subdivisions, channel network, and other derived parameters to those obtained by traditional evaluation

Assessing the Performance of

Automated Watershed Segmentation

from Digital Elevation Models

By Jurgen Garbrecht and Lawrence W Martz

INTRODUCTION

Watershed segmentation and channel network definition is often required in distributed hydro-logic modeling Manual segmentation from maps is a tedious, time-consuming, and subjective task, particularly for large watersheds The automated watershed segmentation and extraction of channel network and subwatershed properties from raster elevation data represents a convenient and rapid way to parameterize a watershed The increasing availability of DEM coverage for many areas of the United States makes this automated watershed segmentation and characterization a promising approach for a wide range of hydrologic investigations However, assessment of the generated watershed data beyond the usual visual inspection is required to develop confidence in the automated approach

Early research on automated landscape analysis focused on algorithm development and treat-ment of unique situations, such as depressions, flat areas, or the connectivity of the network DEM processing models which identify upward concave areas (Pueker and Douglas, 1975; Jenson, 1985; Band, 1986) often produce discontinuous network segments that must subsequently be con-nected (O’Callaghan and Mark, 1984) and may require additional adjustments to produce a rea-sonable pattern (Douglas, 1986) Other models rely on flow routing concepts In this approach, the steepest downslope direction defines the flow paths (Jenson and Domingue, 1988; Martz and de Jong, 1988; Morris and Heerdegen, 1988) For either approach, problems arise in low relief terrain when the vertical resolution of the DEM is insufficient to identify either upward concave areas or

a downslope flow direction Little work has been done to extend DEM processing methods to low relief terrain, such as found in the central plains of the U.S

This chapter investigates the performance of automated watershed segmentation, channel net-work identification, and subwatershed definition in low relief terrain for an 84 km2watershed The landscape analysis computer program chosen from this study is the Digital Elevation Drainage Net-work Model DEDNM (Martz and Garbrecht, 1992, 2001) Program DEDNM is the main compo-nent of a larger landscape analysis program called TOPAZ (TOpographic PArameteriZation) (Garbrecht and Martz, 1999, 2000) The approach used by program DEDNM is similar to that of other DEM processing models that are based on flow routing concepts, but it includes enhance-ments for processing low relief landscapes where the rate of elevation change may be only a few meters per kilometer over large areas The program is designed for the resolution of the USGS 7.5′ DEMs (1 meter or 1 foot elevation increments for 30 m grid cells) and targets parameters that are of

hydrologic interest Even though program DEDNM was designed primarily for hydrologic and water resources investigations, it is equally applicable to address a variety of geomorphological, en-vironmental, and remote sensing applications The performance of program DEDNM is evaluated

by comparing the generated watershed subdivisions, channel network, and other derived parameters

to those obtained by traditional evaluation methods using USGS 7.5′ minute topographic maps

THE APPLICATION WATERSHED

Bills Creek watershed drains an 84 km2area of the Little Washita River watershed, a USDA-ARS experimental watershed in southwestern Oklahoma The terrain consists of gently rolling hills, and land use is predominantly rangeland with some cultivated areas The source and the out-let of Bills Creek are at 420 m and 336 m above MSL, respectively With a longitudinal distance

of 16.5 km, the average main channel slope is about 0.005 m/m Channel slopes as low as 0.003 m/m and flat flood plain areas extending over several hectares are not uncommon toward the out-let of Bills Creek watershed

USGS 7.5′ DEM coverages from the geographic area under consideration are not available, and

a previously generated DEM of the ARS experimental watershed is used to conduct this study The DEM was generated in 1987 by the NASA Stennis Laboratory, Slidell, Louisiana, from USGS 7.5′ topographic maps Elevation data are given for 30 m grid cells in 0.91 m (3-foot) increments These DEM characteristics are similar to the USGS 7.5′ DEM coverages which have a 30 m grid spacing and a 1 m vertical resolution A three-dimensional, vertically enhanced representation of Bills Creek watershed is shown in Figure 2.1 (see color section) With a vertical resolution of 0.91

m, parts of the gently sloping floodplains appear as flat areas and include a number of apparent de-pressions or pits which are artifacts of the DEM Such spurious dede-pressions are often encountered

in low resolution DEMs of terrain with limited relief and usually arise from input errors, interpo-lation procedures, elevation rounding and limited DEM resolution These produce under- and overestimation of elevation values of individual or groups of DEM cells which then can result in spurious depressions The extent of the flat areas and depressions in Bills Creek watershed are il-lustrated in Figure 2.2 (see color section) They can be seen to be closely associated with channel bottom lands and drainage divides The low channel slopes, flat areas, and spurious depressions in the DEM of Bills Creek watershed make it an appropriate low relief application for evaluating wa-tershed segmentation and channel network generation capabilities of program DEDNM

The channel network, defined by the blue-line channels on the USGS 7.5′ topographic maps (Figure 2.3a), is the map-based network against which the generated network is assessed It should

be recognized that a channel network is a dynamic drainage feature that can change in time as a function of climate, land use, and other land surface parameters Thus, any channel network defined from topography alone is the reflection of past runoff and erosional activities and may not neces-sarily be representative of current runoff conditions Other methods such as contour crenulation or slope analysis can also be used to define the channel network, but they also have limitations The blue-line method (Morisawa, 1957) was selected because it is the simplest and one that most read-ers are familiar with, and despite its limitations, it is a standard that is readily available for most wa-tersheds in the United States The map-based network of Bills Creek watershed is of 5th Strahler order (Strahler, 1957) and of magnitude 180 following Shreve’s (Shreve, 1967) classification

BOUNDARY CONDITIONS FOR WATERSHED SEGMENTATION

Program DEDNM requires that two channel network parameters be specified for the automated watershed segmentation: the critical source area and the minimum source channel length The

crit-ical source area is the drainage area required to support a permanent channel Its value is related to soil characteristics, vegetation cover, climatic conditions, and terrain slope It also varies with map scale when maps are used as the basis for network definition (Scheidegger, 1966) For the present application, the critical source area was determined from the USGS 7.5′ topographic maps and was about 9 ha The minimum channel length for source channels is a parameter that is necessary

to control the identification of very short channels that satisfy the critical source area criterion, but have no real significance at the scale of the USGS 7.5′ topographic maps Therefore, a threshold length below which first order channels are not generated is specified This threshold length for Bills Creek was determined from the USGS topographic maps to be 130 m With only these two input channel network parameters, the DEM can be processed by program DEDNM, and a seg-mented watershed, and channel network and subcatchment parameters can be extracted from the DEM As part of this processing, the spurious depressions in the DEM are first removed by raising the elevation of the cells within the depressions to the elevation of the lowest outlet cell on the out-side edge of the depression The watershed segmentation and channel network and subcatchment parameter extraction is then performed on this depression-free DEM

RESULTS

Program DEDNM generates raster maps of the channel network, subwatersheds, and other drainage parameters The watershed segmentation is represented by the drainage boundaries of the subcatchments and the channel network As examples of generated raster maps, the raster map of elevation contours, flow vectors, and subwatersheds boundaries, including the generated channel network, are displayed in Figures 2.4, 2.5, and 2.6, respectively (see color section) Program DEDNM also produces attribute tables that can be used for subwatershed parameterization, dis-tributed surface runoff modeling, or other purposes In the following discussion, network and

sub-Figures 2.3a and 2.3b Channel network of Bills Creek watershed: (a) blue-line network from USGS 7.5′

topographic maps; (b) generated network by DEDNM.

watershed raster and attribute data are compared to values derived from USGS 7.5′ topographic maps by traditional methods

Evaluation by Visual Appearance

The channel network generated by program DEDNM (Figure 2.3b) is similar to the blue-line map-based network derived from the USGS 7.5′ topographic maps (Figure 2.3a) Differences are apparent, particularly at the level of source channels Some source channels are generated at loca-tions where there are none on the USGS maps, although contour crenulation suggests the presence

of a channel in most of these cases The reverse is also true; source channels with small upstream source area are shown on the USGS maps, but are not generated by program DEDNM Other source channels differ by their length The primary reasons for these differences are the ambiguity

in the definition of a source of a channel, natural variation in channel characteristics within the network, and the use of only two parameters to generate the entire network The two-parameter approach used by program DEDNM (critical source area and minimum source channel length) can reproduce average channel properties, but cannot account for random or spatial variability of these properties within the watershed When such spatial variability is important, the user must either apply the program to each homogeneous subarea separately, or introduce a variable channel main-tenance constant and minimum source channel length

The differences between the two networks diminish for higher channel magnitude or Strahler order This increasing similarity is demonstrated for third and higher Strahler order channels (Fig-ure 2.7a), and for fifth and higher Shreve magnitude channels (Fig(Fig-ure 2.7b) In both cases, the main channel, the East and West Fork, and several larger tributaries are well reproduced The few

Figures 2.7a and 2.7 b Generated and measured channel networks for (a) channels of third and higher Strahler order; (b) channels of fifth and higher Shreve magnitude.

remaining differences are channels that exist in both networks, but are not represented because

they did not meet the order or magnitude selection criterium used to generate the illustrations

Comparison of Selected Channel Network Parameters

The selected parameters used to compare the generated and map-based channel networks are

listed in Table 2.1 and represent the topologic, geometric, and hydrographic characteristics of the

networks The parameter values are representative of the average network characteristics of the

entire network and do not reflect spatial variations within the watershed As indicated earlier,

spa-tial variability can be accounted for by either evaluating each homogeneous subarea separately, or

introducing a variable channel maintenance constant and minimum source channel length

The watershed Strahler order and the Shreve magnitude are nearly identical for the generated

and map-based channel networks The number of channel links and the link lengths are

repro-duced with a mean difference of less than 2%

The channel slope values shown in Table 2.1 are length weighted values; i.e., long channels are

assumed to be more representative of average network conditions The mean slope of the

gener-ated channel links is about 23% lower than the slope derived from the map-based network For

in-terior links the difference is 10%, whereas for exin-terior links it is 27%, suggesting that exin-terior

links are the primary reason for the discrepancy Exterior links are the same as first order channels

or source channels

Table 2.1 Selected Network Parameters for Generated and Map-Based Networks and Corresponding

Deviations

Channel length [m]

Channel slope [m/m]

Drainage area [ha]

The ambiguous and subjective definition of source channels, natural variations in source chan-nel characteristics, the coarse vertical resolution of the DEM, and the use of only two parameters

to generate the entire network are believed to lead to the observed differences A more representa-tive comparison of channel slope is achieved by computing the slope over long channel stretches, such as for the East and West Fork of Bills Creek For these two channel stretches, the mean slope from source to watershed outlet is reproduced with less than 10% discrepancy This illustrates that even though large differences in slope may exist for individual channel links, the average slope over long channel stretches is reasonably approximated by program DEDNM

The total watershed area and mean direct drainage area feeding channel sources and channel links are reproduced within 4% Finally, the drainage density of the generated network is within 5% of that of the validation network

Comparison of Channel Network Composition

The channel network composition is quantified by the bifurcation, length, slope, and upstream area ratios, as defined by Horton (1945) and Schumm (1956) These ratios define the rate of change of a variable with channel order and incorporate a relative measure of the variable magni-tude for each order The channel network composition analysis complements the previous analysis

of selected network parameters because it measures change of parameter value with channel order The average value of the four ratios for the generated network is within 4% of that obtained for the map-based network (Table 2.2) The slope and relative position of the regression lines in Fig-ures 2.8a and 2.8b graphically represent the ratios In all four cases, the regression lines are quite similar in slope and position This close agreement shows that the general character of a system of channels is reproduced by the automated network extraction Variations in parameter values for in-dividual Strahler orders are primarily the result of the stochastic nature of the Strahler ordering sys-tem As previously reported by Gregory and Walling (1973), and Scheidegger (1966 and 1970), the stochastic aspect of Strahler’s ordering sometimes result in different order values for corresponding channel segments in two very similar networks This is why direct comparisons of parameter values for individual Strahler order are not necessarily conclusive and are not performed here

Table 2.2 Ratios of Channel Network Composition

Generated Validation Deviation

DISCUSSION AND CONCLUSIONS

In this chapter, the channel network generated by program DEDNM for an 84 km2low relief watershed is compared to the one defined by the blue-line method on the USGS 7.5′ topographic maps The selected watershed included flat areas and depressions along the valley bottom, and flat areas near drainage divides The depressions are usually artifacts of the DEM and have been re-moved by raising the elevation of the cells within the depression to the elevation of the lowest out-let cell on the outside edge of the depression The watershed segmentation and channel network and subcatchment properties are generated from this depression-free DEM

The visual appearances of the generated and map-based channel networks are very similar, and

so are the channel network composition parameters which display an average discrepancy of less than 5% Other selected network parameters that describe the channel network also show good correspondence For example, the generated watershed Strahler order, the Shreve magnitude, the number of channel links, the channel link length, the drainage area, and the drainage density are within less than 5% of those of the map-based channel network The average discrepancy for all parameters used in this study is also less than 5% The largest differences are found for channel slope The reason for the larger discrepancies in longitudinal channel slope values is primarily the ambiguous definition of first order channels in the map-based channel network, the coarse DEM resolution, and natural variations in channel characteristics within the network which cannot be re-produced by the two-parameter model of program DEDNM Spatial differences between gener-ated and map-based channel networks occur because channel generation criteria are applied uniformly to the entire network and, therefore, cannot reproduce spatial variability within the channel network For investigations where spatial variation is important, program DEDNM should

be applied to subareas that are homogeneous In general, the close agreement between the various parameters describing the overall channel network and subwatershed characteristics demonstrates the ability of program DEDNM to overcome the problems associated with ill-defined drainage boundaries and indeterminate flow paths in low relief terrain

Based on the experience of this application, further improvements have been introduced into program DEDNM These include the capability to reproduce spatial variability in the generated channel network and subcatchments within the watershed and to treat spurious depressions by a combination of breaching and filling which is more appropriate given that spurious depressions arise from elevation over- and underestimation These improvements are covered in a separate chapter of this book Program DEDNM with these improvements and other additions are available

in a software package called TOPAZ (TOpographic PArameteriZation) (Garbrecht and Martz, 1997)

REFERENCES

Band, L E., 1986 Topographic partition of watersheds with digital elevation models Water Re-sources Research, 22(1):15–24.

Figures 2.8a and 2.8b Channel network composition: (a) length and area ratios, R L and R A ; (b) bifurcation and slope ratios, R N and R S .

Douglas, D H., 1986 Experiments to locate ridges and channels to create a new type of Digital

Elevation Model Cartographica 23(4):29–61.

Garbrecht, J., and L W Martz, 1999 TOPAZ: An Automated Digital Landscape Analysis Tool for Topographic Evaluation, Drainage Identification, Watershed Segmentation and Subcatchment Parameterization; TOPAZ Overview U.S Department of Agriculture, Agricultural Research Service, Grazinglands Research Laboratory, El Reno, OK, ARS Publication No GRL 99–1, 26 pp

Garbrecht, J., and L W Martz, 2000 TOPAZ: An Automated Digital Landscape Analytical Tool for Topographic Evaluation, Drainage Identification, Watershed Segmentation and Subcatch-ment Parameterization: TOPAZ User Manual U.S DepartSubcatch-ment of Agriculture, Agricultural Re-search Service, Grazinglands ReRe-search Laboratory, El Reno, OK, ARS Pub No GRL 2-00, 144 pp

Gregory, K J., and D E Walling, 1973 Drainage Basin Form and Process: A Geomorphologic Approach John Wiley and Sons, New York, NY.

Horton, R E., 1945 Erosional development of streams and their drainage basins; Hydrophysical

approach to quantitative morphology Bul Geol Soc Am., 56:275–370.

Jenson, S K., 1985 Automated Derivation of Hydrologic Basin Characteristics from Digital

Ele-vation Data Proc Auto-Carto 7, Digital Repres of Spatial Knowledge, Washington, D.C., pp.

301–310

Jenson, S K., and J Q Domingue, 1988 Extracting topographic structures from digital elevation

data for Geographic Information System analysis Photogrammetric Engineering and Remote Sensing, 54(11):1593–1600.

Martz, L W., and E DeJong, 1988 CATCH: A FORTRAN program for measuring catchment area

from Digital Elevation Models Computers and Geosciences, 14(5):627–640.

Martz, L W., and J Garbrecht, 1992 Numerical definition of drainage network and subcatchment

areas from Digital Elevation Models Computers and Geosciences, 18(6):747–761.

Martz, L W., and J Garbrecht, 2001 Channel network delineation and watershed segmentation in

the TOPAZ digital landscape analysis system In Lyon, J., Ed., 2001, GIS for Water Resources and Watershed Management Ann Arbor Press, Chelsea, MI.

Morisawa, M., 1957 Accuracy of determination of stream lengths from topographic maps Trans.

Am Geo Un., 38(1):86–88.

Morris, D G., and R G Heerdegen, 1988 Automatically derived catchment boundary and

chan-nel networks and their hydrological applications Geomorphology, 1(2):131–141.

O’Callaghan, J F., and D M Mark, 1984 The extraction of drainage networks from digital

ele-vation data Computer Vision Graphics and Image Processing, 28:323–344.

Peucker, T K., and D H Douglas, 1975 Detection of surface specific points by local parallel

pro-cessing of discrete terrain elevation data Computer Vision Graphics and Image Propro-cessing,

4(3):375–387

Scheidegger, A E., 1966 Effect of map scale on stream orders Bul Int Assoc Sci Hydrol., XI, 3 Scheidegger, A E., 1970 Theoretical Geomorphology Second Revised Edition, Springer Verlag,

New York, NY

Shreve, R L., 1967 Infinite topologically random channel networks J Geol., 75:178–186.

Schumm, S A., 1956 Evolution of drainage systems and slopes in badlands at Perth Amboy, New

Jersey Bul Geol Soc Am., 67:595–646.

Strahler, A N., 1957 Quantitative analysis of watershed geomorphology Trans Am Geophys Union, 38(6):913–920.