Storm Runoff Computation Using Spatially Distributed Terrain Parameters

by Francisco Olivera and David R. Maidment University of Texas at Austin Center for Research in Water Resources

1. Introduction
In rainfall-runoff computation, not only is the generation of excess precipitation spatially distributed but also the precipitation itself, which has been a limitation for the use of the classic unit hydrograph model for years. The theory presented in this paper is an attempt to generalize the unit hydrograph method for runoff response, and to do so on a spatially distributed basis in which the runoff responses from subareas of the watershed are considered separately instead of being spatially averaged. Although the theory of linear routing systems presented in this article is not bound to raster representations of the study area, the model proposed here is based on grid data structures. A grid data structure is a discrete representation of the terrain based on identical square cells arranged in rows and columns. Grids are used to describe spatially distributed terrain parameters (i.e. elevation, land use, impervious cover, etc.), and one grid is necessary per parameter that is to be represented. The density of grid cells should be large enough to resemble a continuous character of the terrain. Starting from the digital elevation model (DEM), hydrologic features of the terrain (i.e. flow direction, flow accumulation, flow length, stream-network, and drainage areas) can be determined using standard functions included in commercially available geographic information system software that operates on raster terrain data. At present, DEMs are available with a resolution of 3 arcseconds (approximately 90 m) for the United States, and 30 arc-seconds (approximately 1 Km) for the entire earth, etc. Since in the case of water draining under gravity a single downstream cell can be defined for each DEM cell, a unique connection from each cell to the watershed outlet can be determined. This process produces a cell-network, with the shape of a spanning tree, that represents the watershed flow system. Flow routing consists of tracking the water throughout the cell-network. For this purpose, a two-parameter response function is determined for each cell, in which the parameters are related to flow time (flow velocity) and to shear effects (dispersion) in the cell. Flow-path response functions are calculated by convoluting the responses of the cells located within the reach. Finally, the watershed response is obtained as the sum of the cell responses to a spatially distributed precipitation excess.

A deconvolution algorithm is used to estimate the precipitation excess from flow records instead of from precipitation records. This algorithm consists of deconvolving an observed hydrograph by an estimated watershed response function (unit hydrograph) to obtain the precipitation excess. The spatial distribution of the precipitation excess is assumed to be proportional to the runoff coefficient.

2. Literature Review
Pilgrim (1976) carried out an experimental study consisting in tracing flood runoff from specific points of a 0.39 Km2 watershed, near Sydney, Australia, and measuring the travel time of the labeled particles to the outlet. A conclusion of his study is that "at medium to high flows the travel times and average velocities become almost constant, indicating that linearity is approximated at this range of flows". Additionally, for a watershed subdivided into non-overlapping subareas, linearity of the routing system implies that the overall watershed response is equal to the sum of the responses of its subareas, which is an important insight in dealing with spatial variability of the watershed. An significant attempt to linking the geomorphological characteristics with the hydrologic response of a watershed is given by Rodriguez-Iturbe and Valdes (1979). In their paper, Hortons empirical laws, i.e. law of stream numbers, lengths and areas, are used to describe the geomorphology of the system. Mesa and Mifflin (1986), Naden (1992) and Troch et al. (1994) present similar methodologies to account for spatial variability when determining the watershed response. The catchment response is calculated as the convolution of a network response and a hillslope response. The network response is calculated as the solution of the advection-dispersion equation, weighted according to the width function of the network. However, the researchers present no physically-based methodology to determine the hillslope response. An interesting approach to model the fast and slow responses of a catchment is presented by Littlewood and Jakeman (1992, 1994). In their model, the watershed is idealized as two linear storage systems in parallel, representing the surface and the subsurface water systems. The surface system is faster and affects mainly the raising limb of the resulting hydrograph, while the subsurface system is slow and determines the tail of the response. Geographic Information Systems (GIS) are tools that allow one to jump from lumped to spatially distributed hydrologic models. The border between lumped and distributed models is not sharp, and there are pre-GIS attempts to deal with spatially distributed terrain attributes. For example, the Hydrologic Engineering Center (HEC) flood model HEC-1, well known as a lumped model, allows the user to subdivide the watershed into smaller sub-basins for analysis purposes, and route their corresponding responses to the watershed outlet. In this case, the concept of purely lumped model does not apply, although it cannot be considered

a fully spatially distributed model either. It is therefore advisable to keep in mind the extent to which a given model is lumped or distributed. Grid-based GIS appears to be a very suitable tool for hydrologic modeling, mainly because "raster systems have been used for digital image processing for decades and a mature understanding and technology has been created for that task" (Maidment 1992 a). The ESRI Arc/Info-GRID system as well as the U.S. Army Corps of Engineers GRASS system, work on grid data structures. Grid systems have proven to be ideal for modeling gravity driven flow, because a characteristic of this type of flow is that flow-directions depend entirely on topography and not on any time dependent variable. This characteristic is what makes gravity driven flow easy to be modeled in a grid environment and, consequently, grid systems include hydrologic functions as part of their capabilities. At present, hydrologic functions, available in GRID and GRASS, allow one to determine flow direction and drainage area at any location, stream networks, watershed delineation, and others (Maidment 1992 a). Recently, there have been attempts to take advantage of GIS capabilities for runoff and non-point source pollution modeling. Vieux (1991) presents a review of water quantity and quality modeling using GIS and, as an application example, employs the kinematic wave method to an overland flow problem. GIS is used to process the spatially variable terrain and the finite elements method (FEM) to solve the mathematics. Maidment (1992 a, 1992 b, 1993) presents a grid-based methodology for determining a spatially distributed unit hydrograph that assumes a time-invariant flow velocity field. According to him, the velocity timeinvariance is a requirement for the existence of a unit hydrograph with a constant time base and relative shape. This concept is also explained in this article, in the light of the conditions for linearity of a routing system. In Maidment's articles, from a constant velocity grid, a flow time grid is obtained and subsequently the isochrone curves and the time-area diagram. The unit hydrograph is obtained as the incremental areas of the time-area diagram, assuming a pure translation flow process. A more elaborate flow process, accounting for both translation and storage effects, is presented by Maidment et al. (1996 a). In their paper, the watershed response is calculated as the sum of the responses of each individual grid-cell, which is determined as a combined process of channel flow (translation process) followed by a linear reservoir routing (spreading process). Olivera et al. (1995) and Olivera and Maidment (1996 a) present a grid-based, unsteady-flow, linear approach that uses the diffusion wave method to model storm runoff and constituent transport. In these articles, the routing from a certain location to the outlet is calculated by convoluting the responses of the grid-cells of the drainagepath. Sensitivity of model results to the spatial resolution of the data has been addressed by Vieux (1993), who discusses how the grid-cell size affects the terrain slope and flow-path length, and, in turn, the surface runoff. Vieux and Needham (1993)

conclude that increasing the cell size shortens the streams length and increases the sediment yield. Attempting to account for spatial variability of the terrain in storm runoff modeling, researchers have taken either of the following paths: (1) partitioning the hydrologic system into subsystems and applying lumped models to each of them, or (2) developing GIS interfaces to generate input files for other lumped models, and display the results in the form of a map. In both cases, an improvement with respect to the traditional fully lumped models has been accomplished; however, these kind of solutions can not be considered spatially distributed. In this research, storm runoff is modeled within GIS, redefining the use of GIS by using it as a modeling tool itself and not only as a link between the heterogeneous terrain and an existing non-GIS model.

3. Methodology
For a spatially uniform hydrologic system, the classic unit hydrograph model states that

(1) where t [T] is time, Q(t) [L3T-1] is the flow at the watershed outlet, AW [L2] is the watershed area, I(t) [LT-1] is the excess precipitation, and U(t) [T-1] is the watershed unit hydrograph. Likewise, for a spatially distributed linear system subdivided into uniform non-overlapping subareas, this equation takes the form of (Maidment et al. 1996)

(2) where NW is the number of subareas, Ai [L2] is the area of subarea i, Ii(t) [LT-1] is the excess precipitation in subarea i (subarea input), and Ui(t) [T-1] is the response at the watershed outlet yield by a unit instantaneous input in subarea i. Notice that it is because of the additivity property that characterizes linear systems, that the overall watershed response can be calculated as the sum of the subarea responses. From the physical point of view, this summation implies that the flow of a subarea input to the watershed outlet is not affected by the flow of the other subarea inputs, and that all inputs can be routed simultaneously yet independently. The use of equation (2) requires for each subarea the excess precipitation Ii(t), and the response function Ui(t). In this study, the subareas are taken as small square cells that resemble the continuous character of the landscape (see Figure 1), and, because the number of

cells into which the watershed is subdivided defines how well the spatial variability of the system is captured, the accuracy of the model depends on the resolution of the discretization. Once the system has been discretized into cells, a flow-path is defined as the sequence of cells that connect a specific cell with another cell located downstream (see Figure 2).

Figure 1: Subdivision of the watershed into grid cells. Although for presentation purposes the watershed in the figure has been subdivided only into 24 cells, a large number of cells - in the order of thousands or more - is necessary to capture the spatial variability of the system.

Figure 2: Cell-base representation of a flow-path.

The methodology consists of: (1) calculating the overall watershed response as the weighted sum of the cell responses, where the weight is given by the excess precipitation (see equation (2)); and (2) estimating the spatially distributed excess precipitation by deconvolving the watershed flow records by the watershed response (see equation (1)), and distributing this lumped excess precipitation based on the terrain physical characteristics. It becomes clear that both processes are strongly related, and that one cannot be considered without the effect of the other. Routing the excess precipitation from the terrain to the watershed outlet is covered in Flow Routing, while estimating the volume and spatial distribution of the excess precipitation in Excess Precipitation. 3.1. Flow Routing The response at the watershed outlet cell yield by a unit instantaneous input in a cell is called here flow-path response function Ui(t), and consists of two parts: a flow-path redistribution function U'i(t) [L-1] that represents the translation and redistribution processes in the flow-path (lag-time from the cell to the outlet and spreading around the centroid of the mass slug); and a flow-path loss factor Ki(t) [-] that accounts for the losses along the flow-path. Note that, because of how losses are accounted for, the area under the curve U'i(t) vs. t is equal to one, and the values of Ki(t) are less than one (see Figure 3).

Figure 3: The flow-path response function results from the product of the flowpath redistribution function and the flow-path loss factor. Flow-path redistribution functions U'i(t) have to satisfy certain mathematical properties so that if, for example, an input in cell A is routed to cell B and then to the cell C, the result should be the same as if it were routed directly from A to C (see Figure 4).

Figure 4: Flow-path from cell A to cell C.

To understand the implications of this condition, notice that the flow-path redistribution function U'AB(t) is the probability density function of a random variable XAB that represents the time spent by a water particle in the flow-path that runs from cell A to cell B. Accordingly, U'BC(t) and U'AC(t) are the probability density functions of random variables XBC and XAC, respectively. Since the time spent in AC is the sum of the times spent in AB and BC, it follows that XAC = XAB + XBC. In terms of probability density functions, this is expressed as (3) where * stands for convolution integral. Equation (3) implies that the redistribution functions should be self reproducing, i.e., the convolution of two redistribution functions results in a function of the same type. This condition precludes one, for example, from defining the redistribution functions as exponential distributions; in other words, from modeling the watershed as an array of linear reservoirs with one linear reservoir per grid cell, which is a common but erroneous approach. In statistical terms, the type of functions that can be used as redistribution functions are called infinitely divisible distributions. The normal, gamma and first-passage-times distributions are examples of infinitely divisible distributions. From the physical point of view, if the flow-path is assumed to convey onedimensional unsteady flow and the inertial terms in the St. Venant momentum equation are neglected, the flow can be modeled with the diffusion wave equation (Miller and Cunge, 1975, Lettenmaier and Wood 1993). Thus, if lateral inflow is not considered, the flow is represented by

(4) where x [L] is the distance along the flow-path, t [T] is the time, qi(t) [L3T-1] is the flow at any time and point of the flow-path, Ci [LT-1] is the kinematic wave celerity, and Di [L2T-1] is a dispersion coefficient. For a unit-impulse input, the solution for qi of equation (4) at the flow-path outlet is the flow-path redistribution function U'i(t), and it results in a first-passage-times distribution (Nauman 1981):

(5) where Ti = Li / Ci is the mean value of the distribution (the lag time in the flowpath), i = Ci Li / Di represents the spreading around the mean of the distribution (the shear and storage effects on the flow), and Li [L] is the flow-path length. First-passage-times distributions apply to systems bounded by a transmitting

barrier upstream (open boundary) and an adsorbing barrier downstream (close boundary). First-passage-times distributions are in accordance with what other researchers have proposed to model the time spent by water in hydrologic systems (Mesa and Mifflin 1986, Naden 1992, Troch et al. 1994). Likewise, it has been shown that other infinitely-divisible two-parameter distributions, say normal or gamma, do not differ significantly from the first-passage-times distribution with the same first and second moments (Olivera 1996 b). Olivera (1996 b) has also observed that three-parameter distributions tend to overestimate the importance of the tails with respect to the central part of the distribution. Extending the concept of self-reproducing flow-path redistribution functions to the cell level, allows one to model the flow based on scale-independent terrain parameters. Since the time spent in a flow-path is equal to the sum of the time spent on each its constituting cells, i.e., Xi = x1 + x2 + ... + xN where Xi is a random variable that represents the time spent in the flow-path and x1, x2, ...and xN are random variables that represent the time spent in each of the N cells that form the flow-path, it follows (6) where U'i(t), u'1(t), u'2(t), ... u'N(t) [T-1] are the probability density functions of Xi, x1, x2, ... and xN respectively. Moreover, because Ui(t) is a first-passage-times distribution, u'1(t), u'2(t), ... and u'N(t) are also first-passage-time distributions that can be expressed as

(7) where j refers to the cell of the flow-path, vj [LT-1] is the flow velocity, dj [L2T-1] is the dispersion coefficient (shear and storage effects), tj [T] is the expected flow time through the cell and lj = vj tj. Because the cell flow length lj is known, the only two parameters needed to define u'j(t) are the flow velocity vj and the dispersion coefficient dj. In some cases, it is preferable to define the dimensionless Peclet number vjlj/dj - instead of the dispersion coefficient dj - to describe the shear and storage effects in the cell. However, it should be noted that, because it involves the flow length in its definition, the Peclet number is a scale dependent parameter. The connection between the flow-path redistribution function and the cell redistribution functions is given by equation (6). However, the use of this equation implies as many convolution integrations as cells are there in the watershed. Depending on the hardware available, this process might be extremely demanding and time consuming. A good approximation to U'i(t) - whose error falls within the limits of the uncertainty of the model parameters - can be obtained based on the fact that (DeGroot 1986)

(8) where E refers to expected value (first moment) and Var to variance (second moment around the mean). According to this method, the approximate solution for U'i(t) has the same first and second moments as the solution obtained with equation (6). By equalizing the first and second moments of U'i(t) given by equation (5) to the sum of the moments of the u'j(t) given by equation (7), relations between Ti and i, and vj and dj are determined as

(9) and

( 10 ) The main advantage of this approach is that it can be applied automatically by using standard functions like the weighted flow length function - included in commercially available geographic information systems software that operates on raster terrain data. Water losses in linear systems are represented by a first-order loss term in the mass balance equation

( 11 ) where i [T-1] is a flow-path loss coefficient. The solution of equation (11) is given by ( 12 )

where U'i(t) - the flow-path redistribution function - is equivalent to the flow-path response function when losses are neglected, and exp(- i t) represents the losses in the flow-path. The flow-path loss factor Ki(t) is, therefore, given by ( 13 ) Similarly, for a cell of a flow-path, it can be demonstrated that ( 14 ) where kj(t) [-] is the cell loss factor and j [T-1] is a cell loss coefficient. For small losses, the cell loss factors can be approximated to the constant value kj = exp(- j tj), and the flow-path loss factor to the product of the cell loss factors

( 15 ) which, considering that tj = lj / vj, is equal to

( 16 ) Finally, the flow-path response function can be expressed as

( 17 ) where Ti, i and Ki are given in equations (9), (10) and (16) as functions of the flow velocity vj, the dispersion coefficient Dj and the loss coefficient j. Note that if the cell inter-connectivity and the grids of vj, Dj and j are defined, the terrain would be fully described for flow routing purposes. 3.2. Excess Precipitation One of the advantages of the theory of linear routing systems is that it can handle spatially distributed inputs, letting the excess precipitation vary according to the terrain physical characteristics, say land use, soil type or topography. Standard engineering practice estimates excess precipitation based on soil-water balance. Willmott et al. (1985), for instance, have developed the WATBUG

Fortran program that simulates the soil-water balance based on local temperature, precipitation and soil water-holding capacity. Although the soil-water balance is a physically based approach, it has been observed that it is a complicated process that should account for a large number of parameters, that it is very sensitive to the data available, and that it produces results that have to be interpreted with extreme caution. Other simple excess precipitation models such as the Soil Conservation Service (SCS) curve number method, or just the product of a runoff coefficient by the precipitation, are examples of attempts of solving the problem. However, estimating the correct excess precipitation is still far from being achieved. In the following, a deconvolution methodology for determining excess precipitation from flow instead of from precipitation records is presented. Given the flow at a specific station, and a spatially-distributed parameter that describes the terrain tendency to generate runoff, the spatially-distributed excess precipitation and flow parameters are calculated. The method consists of deconvolving the observed direct runoff by the watershed unit response function. This unit response is estimated from the flow records, and considers the watershed as a lumped system. Spatial variability of the terrain is considered later in the process. The relation between direct runoff and excess precipitation is given by ( 18 ) where Q(t) [L3T-1] is the direct runoff, A [L2] is the watershed area, r(t) [T-1] is the estimated watershed unit response, and Pe(t) [LT-1] is the excess precipitation. Determining the excess precipitation consists of solving equation (18) for Pe. The excess precipitation is calculated at discrete time steps by trial and error, guessing values of Pe for each time step and then verifying if equation (18) is satisfied. An optimization software helped in the process of determining the excess precipitation values. Since the flow and unit response function are considered for the watershed as a whole, the excess precipitation obtained by deconvolution is a lumped type of result that should be distributed according to the local hydrologic characteristics of the terrain. It is assumed that the amount of excess precipitation produced by each cell is a function of the runoff coefficient, a well known hydrologic parameter that can be estimated from tables available in the literature (Chow et al. 1988, Browne 1990, Pilgrim and Cordery 1993). A connection between excess precipitation in a cell and runoff coefficient is a critical assumption that allows one to use a simple model without going through a more complicated - but not necessarily more reliable - soil-water balance. In our model, excess precipitation in a cell Ii is assumed to be proportional to the runoff coefficient ci minus a uniformly distributed abstractions parameter (i.e., Ii = ci - or 0 whichever is greater). This abstractions parameter constitutes a threshold value, and addresses the fact that low-developed areas might yield no runoff at all. The

abstractions parameter may change from event to event depending on rainfall intensity, soil infiltration capacity and antecedent moisture condition. Values of ci - for each cell are calculated and used as an excess precipitation scale factor. The excess precipitation generated in a cell Ii is given by ( 19 ) where

( 20 ) Aj is the area of cell j, and subscript i refers to the cell where the excess precipitation is being calculated. Note that the values of i have an average value of one. To calculate the spatially-distributed flow parameters (flow velocity vi and dispersion coefficient Di), the watershed unit response, used for the deconvolution in equation (18), is equalized to the weighted sum of the flow-path responses

( 21 ) in which the flow-path responses Ui(t) depend on the flow parameters. In equation (21), the left hand side represents the watershed unit response, while the right hand side is the sum of the flow-path responses corrected by a factor that accounts for the cell area and for its tendency to generate excess precipitation. The values of the flow parameters are tuned until equation (21) is satisfied. The estimated excess precipitation and flow parameters can then be extrapolated to other areas, if the same hydrologic behavior is assumed. This assumption, though, might be questionable when the areas used for calibration and application are dissimilar. The values can also be used to estimate flow hydrographs at other locations within the watershed, where hydrologic dissimilarity is less likely to occur.

4. Case study: Waller Creek in Austin, Texas

Waller Creek is a 14.8 Km2 (3662 acres) watershed located within the urban core of the City of Austin, Texas. Two flow gauging stations, set up by the US Geological Survey (USGS), are located at 23rd and 38th Streets and have drainage areas of 10.7 Km2 (2,643 acres) and 5.7 Km2 (1,416 acres) respectively. The model was calibrated with flow records of the station at 23rd Street and applied to

Waller Creek at 38th Street for comparison with observed flows. The period of analysis ranged from October 14, 1994, 7:45 p.m. to October 17, 1994, 6:45 p.m.. A time step of 15 minutes was used and a total of 284 time intervals were considered. The watershed was delineated using a 30 m digital elevation model (DEM), and comprised approximately 16,500 grid-cells. A map of the drainage-area of the two flow-gauging stations and of the spatial distribution of the runoff coefficient is presented in Figure 5. It can be noted that just upstream of 38th Street there is a large low-developed area that generates little runoff; while just upstream of 23rd Street the area is more developed, yielding much more excess precipitation.

Figure 5: Waller Creek Watershed: Drainage area of the flow gauging stations at 23rd and 38th Streets, creeks, and runoff coefficient distribution. Note that only the gray areas generated runoff for this time period. Additionally, from the flow records, it was noticed that the flow peaked first at 23rd Street and approximately 30 minutes later at 38th Street, which goes against intuition because the peak time did not increase with drainage area. After the direct-runoff/base-flow separation, it was found that, for this time period, 88% of the flow was direct runoff and 12% base flow. Note that, because of the high impervious cover of the urban areas, direct runoff tends to be much more important than base flow during storm events. The methodology presented above was applied in the following steps: STEP 1: The plot - in semi-logarithmic scale - of the flow record of the 23rd Street station showed almost instantaneous peak times and long, straight and parallel recession curves (see Figure 6), which suggests that the watershed responded to all storm events of the period with the same unit response function. This fact confirms that a linear approximation is satisfactory for this hydrologic system, because the response does not change from event to event.

STEP 2: To calculate the flow parameters (flow velocity and dispersion coefficients), the watershed unit response obtained in STEP 1 was reproduced as the aggregation of the flow-path responses according to equation (21). In order to decrease the number of parameters (one flow velocity and one dispersion coefficient per cell), it was assumed that water flows only as overland flow or stream flow, and a single velocity value was assigned to each type of flow; as well, a uniform dispersion coefficient was taken for the entire watershed. Finally, an abstractions coefficient equal to = 0.4 was assumed. From the physical viewpoint, this implies that all cells with runoff coefficient less than 0.4 generate no surplus. The value of = 0.4 was chosen because most cells have runoff coefficient much greater or much less than this value, and by selecting this number it was presupposed that only highly developed areas generated runoff, while lowly developed areas generated no runoff. After these assumptions, the number of model parameters was reduced to three: (1) overland flow velocity, (2) stream flow velocity, and (3) dispersion coefficient. By running an optimization routine, it was found that the flow parameters that produced the best match

between the observed watershed response and the one obtained as the aggregation of the flow-path responses were: overland flow velocity = 0.27 m/s (0.898 fps), stream flow velocity = 0.27 m/s (0.898 fps) and dispersion coefficient = 1,629 m2/s (17,535 ft2/s) (see Figure 7). Note that the optimization routine determined the same value for both velocities, which indicates that in urban areas water flows as fast over impermeable areas as it does in streams.

STEP 3: The watershed unit response, based on the parameters just calculated, was then used to estimate the excess precipitation by deconvolution. Note that the calculated excess precipitation and the observed flow follow the same trend, although the excess precipitation consists of somewhat concentrated pulses, while the flow exhibits long recession curves following short rising limbs (see Figure 8).

STEP 4: Predicted flow in Waller Creek at 38th Street was determined using the excess precipitation calculated in STEP 3, and the model parameters obtained in STEP 2. Figure 9 presents observed flow (labeled Observed) and predicted flow at 38th Street. Three predicted flow series are plotted in this figure: the first one (labeled No abstractions) assumes an abstractions parameter = 0, i.e., cell contributions proportional to the runoff coefficient; the second one (labeled Abstractions = 0.4) assumes an abstractions parameter = 0.4, i.e., cell contributions proportional to the runoff coefficient minus 0.4; and the third one (labeled Proportion) is obtained as the flow at 23rd Street multiplied by the ratio of the two drainage areas. It was interesting to notice that, at least in this case, the No abstractions series and the Proportion series were almost identical, the difference being negligible for practical purposes. With regard to the No abstractions series, it was observed that: (1) predicted values were consistently higher than observed values, yielding a predicted flow volume that was 41% greater than the observed volume; and (2) predicted values followed the trend of the observed values, but shifted approximately 30 minutes (2 time-steps) to the

left. To a great extent, these problems were solved in the Abstractions = 0.4 series. In this case: (1) the flow volume error went down to 4%, (2) the peak times matched and no time-shift was observed, and (3) the recession curves were reproduced well.

The fact that the flow at 38th Street is only 39% of the flow at 23rd Street, instead of 53% as the ratio of the areas, and the fact that the flow peaks first at 23rd Street and 30 minutes later at 38th can be explained in the following way: (1) 38th Street is fed by less developed areas than 23rd Street; (2) as an average, the developed areas that fed 38th Street are farther from the station than those that fed 23rd. This explanation matches the geography of the area, and accounts for the peak shift and runoff volume error. This type of hydrologic behavior evidences the need of accounting for the spatially variability of the hydrologic system.

5. Conclusions

It is known that the spatial variability of the terrain strongly affects storm runoff processes. While representing a watershed by a small number of lumped parameters (i.e. drainage area, channel slope, time of concentration) has the advantage of simplifying the hydrologic modeling, it might miss some specific local processes that affect the overall response of the system and that can not be considered by lumped models. Attempting to account for spatial variability of the terrain in storm runoff modeling, researchers have taken either of the following paths: (1) partitioning the hydrologic system into subsystems and applying lumped models to each of them, or (2) developing GIS user-interfaces to generate input files for, and display the results of, other lumped models. In both cases, an improvement with respect to the traditional fully lumped models has been accomplished, but these kind of solutions can not be considered spatially distributed. In this article, the use of GIS as a modeling tool itself, and not only as a link between the heterogeneous terrain and an existing non-GIS lumped model, has been presented. The model developed here is a generalization of the unit hydrograph model. This generalized version of the unit hydrograph is used to route water in the landscape, provided that the hydrologic system is assumed to be linear. The model also allows the user to consider time- and space-varying rainfall, thus relaxing some of the basic assumptions of the unit hydrograph. The assumption of linearity, though, has not been relaxed by using this approach. GIS appear to be an excellent environment for modeling spatially distributed hydrologic processes, because they have spatial functions in the vector and raster domain (some of them specifically developed for hydrologic purposes) and a database management system, which combined allow one to perform hydrologic modeling and calculations connected to geographic locations. In fact, GIS is able to store and handle more spatially distributed terrain data than can be physically obtained from the field. Thus, when dealing with distributed models, the problem is not necessarily how to develop GIS-based hydrologic models, or how to store and handle the data, but how to get data in an amount that is consistent with the model and hardware/software capabilities. At present, one of the limitations of this type of models is the scarcity of spatially distributed data. With regard to GIS software, Arc/Info-Grid provides the necessary functions and commands to analyze the digital elevation model (DEM) and obtain hydrologic features such as watersheds, drainage areas, and flow lengths. It also provides the weighted flow length function, FLOWLENGTH, that has been used to calculate the first and second moments of the cell-outlet responses automatically, thus performing sequences of convolution integrals - in an approximated way - within the Arc/Info-Grid environment. The importance of accounting for spatial variability of the terrain when modeling storm runoff was evidenced by the case of Waller Creek in Austin, Texas. In this watershed, and according to the data set used here, peak-time did not increase

with drainage area, and it was observed that the flow peaked first at 23rd Street and 30 minutes later did so upstream at 38th Street. This situation reflected that not all the terrain generates the same amount of runoff, and that the relative location of impervious areas in urban watersheds should be considered when attempting to predict flows. Predicted flow for Waller Creek at 38th Street matched reasonable well observed records. Total runoff volume, peak times and recession curves were reproduced well, and although some peak flow values were not matched, the overall tendency was reproduced. This implies that not only the correct amount of water was routed, but also at the correct velocity and with the correct spreading tendency. Although definite conclusions could be drawn only after extensive testing of the methodology, it has become clear that the routing model is an improvement on the currently used unit hydrograph, and that the handling of spatially distributed data by Arc/Info proved to be adequate. However, a more elaborated excess precipitation model might be necessary, especially if the excess precipitation calculated at one watershed is to be exported to other watersheds. 6. References 1. Browne, F.X., Stormwater Management, in Standard Handbook of Environmental Engineering, ed. by R.A. Corbitt, McGraw-Hill Inc., New York, pp. 7.1-7.135, 1990. 2. Chow, V.T, D.R. Maidment and L.W. Mays, Applied Hydrology, McGraw-Hill Inc., New York, 1988. 3. Clark, C.O., Storage and the Unit Hydrograph; Trans. Am. Soc. Civ. Eng., ASCE, Vol 110, pp.1419-1488, 1945. 4. DeGroot M.H., Probability and Statistics, Addison-Wesley Publishing Company, Reading, Mass., 1986. 5. Lettenmaier D. P. & E. F. Wood, Hydrologic Forecast, in Handbook of Hydrology, ed. by D.R. Maidment, McGraw-Hill Inc., New York, 26.126.30, 1993. 6. Levenspiel O., Chemical Reaction Engineering, Wiley, New York, 1972. 7. Littlewood, I.G. and A.J. Jakeman, Characterization of Quick and Slow Streamflow Components by Unit Hydrographs for Single- and Multi-basin Studies, in Methods of Hydrologic Basin Comparison, ed. by M. Robinson, Institute of Hydrology, Report 120, pp. 99-111, 1992. 8. Littlewood, I.G. and A.J. Jakeman, A New Method of Rainfall-Runoff Modelling and its Applications in Catchment Hydrology, in Environmental Modelling, ed. by P. Zannetti, Computational Mechanics Publications, Vol II, pp. 143-171, Southampton, UK., 1994. 9. Maidment, D.R., Grid-based Computation of Runoff: A Preliminary Assessment, Hydrologic Engineering Center, US Army Corps of Engineers, Davis, California, Contract DACW05-92-P-1983, 1992 a.

10. Maidment, D.R., A Grid-Nework Procedure for Hydrologic Modeling, Hydrologic Engineering Center, US Army Corps of Engineers, Davis, California, Contract DACW05-92-P-1983, 1992 b. 11. Maidment, D.R., Developing a Spatially Distributed Unit Hydrograph by Using GIS, in HydroGIS 93, ed. by K. Kovar and H.P. Nachtnebel, Int. Assn. Sci. Hydrol. Publ. No. 211, pp 181-192, 1993. 12. Maidment, D.R., J.F. Olivera, A. Calver, A. Eatherral and W. Fraczek, A Unit Hydrograph Derived From a Spatially Distributed Velocity Field, Hydrologic Processes, Vol 10, No. 6, pp.831-844, John Wiley & Sons, Ltd., 1996 a. 13. Mesa, O.J. and E.R. Mifflin, On the Relative Role of Hillslope and Network Geometry in Hydrologic Response, in Scale Problems in Hydrology, ed. by V. K. Gupta et al., pp.1-17, D. Reidel Publishing Company, 1986. 14. Miller, W.A. and J.A. Cunge, Simplified Equations of Unsteady Flow, in Unsteady Flow in Open Channels, ed. by K. Mahmood and V. Yevjevich, Vol. 1, chapter 5, Water Resources Publications, Fort Collins, CO., 1975. 15. Naden, P.S., Spatial Variability in Flood Estimation for Large Catchments: The Exploitation of Channel Network Structure, Journal of Hydrological Science, 37, 1, 2/1992, pp.53-71, 1992. 16. Nash, J.E. The Form of the Instantaneous Unit Hydrograph, IASH publication No. 45, Vol. 3-4, pp. 114-121, 1957. 17. Nauman, E.B., Residence Time Distributions in Systems Governed by the Dispersion Equation, Chemical Engineering Science Vol. 36 pp.957-966, 1981. 18. Olivera, F., D.R. Maidment and R.J. Charbeneau, Non-Point Source Pollution Analysis with GIS, Proceedings, Spring Meeting, ASCE Texas Section, pp.275-284, April 26-28, Waco, Texas, 1995. 19. Olivera, F., and D.R. Maidment, Runoff Computation Using Spatially Distributed Terrain Parameters, Proceedings, ASCE - North American Water and Environment Congress '96 (NAWEC '96), Anaheim, California, June 22-28, 1996 a. 20. Olivera, F., Doctoral Dissertation, Department of Civil Engineering, University of Texas at Austin, 1996 b. 21. Pilgrim, D.H., Travel Times and Nonlinearity of Flood Runoff From Tracer Measurements on a Small Watershed, Water Resources Research, Vol. 12, No. 3, pp 487-496, June 1976. 22. Pilgrim, D.H., and I. Cordery Flood Runoff, in Handbook of Hydrology, ed. by D.R. Maidment, McGraw-Hill Inc., New York, pp. 9.1-9.42, 1993. 23. Rodriguez-Iturbe, I. and J.B. Valdes, The Geomorphologic Structure of Hydrologic Response, Water Resources Research, Vol. 15, No. 6, pp. 1409-1420, December, 1979. 24. Troch, P.A., J.A. Smith, E.F.Wood and F.P. de Troch, Hydrologic Controls of Large Floods in a Small Basin, Journal of Hydrology, 156, pp. 285-309, 1994.

25. Vieux, B.E., Geographic Information Systems and Non-Point Source Water Quality and Quantity Modeling, Hydrological Processes, Vol. 5, pp. 101-113, 1991. 26. Vieux, B.E., DEM Agregation and Smoothing Effects on Surface Runoff Modeling, Journal of Computing in Civil Engineering, Vol. 7, No. 3, pp. 310-338, July, 1993. 27. Vieux, B.E. and S. Needham, Nonpoint-Pollution Model Sensitivity to Grid-Cell Size, Journal of Water Resources Planning and Management, Vol. 119, No. 2, pp. 141-157, March/April, 1993. 28. Willmott, Cort J., Clinton M. Rowe and Yale Mintz, Climatology of the Terrestrial Seasonal Water Cycle, Journal of Climatology, Vol. 5, pp. 589606, 1985.