flood modeling for complex terrain using gis

Water Resources Management (2005) 19: 605–624DOI: 10.1007/s11269-005-6808-x C© Springer 2005

Flood Modeling for Complex Terrain Using GISand Remote Sensed Information

Y. B. LIU∗ and F. DE SMEDTDepartment of Hydrology and Hydraulic Engineering, Vrije Universieit Brussel, Pleinlaan 2,B-1050 Brussels, Belgium(∗author for correspondence, e-mail: [email protected])

(Received: 11 August 2003; in final form: 11 November 2004)

Abstract. A spatially distributed hydrological model WetSpa (Water and Energy Transfer betweenSoil, Plants and Atmosphere) working on an hourly time scale is presented in this paper. The modelcombines elevation, soil and land use data, and predicts flood hydrograph and the spatial distributionof hydrological characteristics in a watershed. The model is tested on a small catchment in Belgiumfor which topography and soil data are available in GIS form, while the land use and soil cover isobtained from remote sensed images. The resulting calculated discharges compare favorably with thefield measurements. Next a 102-year series of measured hourly precipitation data is processed withthe model and the resulting hydrographs are analyzed statistically to determine the characteristics ofextreme floods. Finally, the simulated extreme peak discharges are compared to the results calculatedwith design storms. Comparison of the two methods shows that the model is capable to predictboth normal and extreme floods. Since the model accounts for spatially distributed hydrological andgeophysical characteristics of the catchment, it is suitable for simulating hydrological processes in acomplex terrain and for predicting the influence of changes in land use on the hydrological behaviorof a river basin.

Key words: flood prediction, GIS, hydrological modeling, remote sensing, runoff

1. Introduction

In applied hydrology, the prediction of peak flow and the simulation of flood hy-drographs in a stream or river is a very complex process, because the hydrologicalvariables vary both in space and time as a function of the meteorological inputs,spatial variability of topography, land use and soil types. Traditional hydrologicallumped models use transfer functions relating statistical properties of rainfall inriver basins to observed runoff and hydrographs. However, these models do notconsider distributed rainfall and basin characteristics, and thus provide little or nospatially distributed information.

In contrast to lumped models, distributed models attempt to account for the spa-tial variability of basin parameters, as well as their physical significance. In floodprediction and rainfall-runoff computation, physically based distributed hydrolog-ical models have become a more feasible approach in recent years. In addition tothe development of improved computational capabilities, Digital Elevation Model

606 Y. B. LIU AND F. DE SMEDT

(DEM), digital data of soil type and land use, as well as the tools of Geographi-cal Information System (GIS), give new possibilities for hydrological research inunderstanding the fundamental physical processes underlying the hydrological cy-cle and the solution of mathematical equations representing those processes. In araster-based hydrological model, the watershed is subdivided into a grid of cells,and model parameters are assigned to each grid cell based on the physical land,soil and vegetation characteristics that exist in that cell. Precipitation and othermeteorological data are then applied over each cell in the watershed, and the runoffis computed and routed along its flow direction to the collecting channel. In such away, distributed models are able to account for the spatial variability of hydrologicalprocesses within a watershed. In addition, the model parameters of this approachare largely physically based and the spatial information of the land, soil, vegetationand precipitation can be captured with much greater detail than the lumped water-shed modeling. Some of these physically based distributed models have obtaineda worldwide recognition, as for instance Topmodel (Beven and Kirkby, 1979) andSHE (Abbott et al., 1986).

Recently, many hydrological models with a flood prediction component usinginformation on topography available from DEM have been developed, whereasmodels like SHE and TOPMODEL were adapted to a new type of data which canbenefit from the GIS techniques (Ewen et al., 2000; Quinn et al., 1991). At thesame time, hydrological models compatible with remotely sensed data and GIShave been developed or updated from their previous version, such as the modelCASC2D (Downer et al., 2002), HYDROTEL (Fortin et al., 2001), and so on.These models are either loosely or tightly coupled with the GIS and remote senseddata.

In this paper, a physically based distributed hydrological model, called WetSpa,is presented, which is tightly coupled with GIS technology and remote sensedinformation. The model takes into account the detailed basin characteristics topredict flood hydrographs and other spatially distributed hydrological variables oncatchment scale. The parameters of the model are derived from DEM, land use andsoil maps in raster format. The model is validated by comparing calculated andobserved hourly discharges for a 16-month period in a small watershed Barebeek,located northeast of Brussels, Belgium, where the topography and soil data areavailable in GIS form, and land use data was obtained from remote sensed images.The utility of the model is demonstrated by forecasting peak discharges resultingfrom an observed 102-year precipitation series. The simulation results are thencompared with the results computed from design storms.

2. The WetSpa Model

The WetSpa model is a grid-based distributed hydrological model for water and en-ergy transfer between soil, plants and atmosphere, which was originally developedby Wang et al. (1997) and adopted for flood prediction on hourly time step by De

FLOOD MODELING FOR COMPLEX TERRAIN 607

Table I. WetSpa model components and their respective parameters

Component Respective model parameters

Interception Interception storage capacity

Depression Depression storage capacity

Surface runoff Potential runoff coefficient, impervious fraction, soil porosity

Interflow Root depth, slope, hydraulic conductivity, scaling factor

Groundwater flow Groundwater flow recession constant

Percolation Hydraulic conductivity, soil porosity, residual soil moisture, soil poresize distribution index

Evapotranspiration Potential evapotranspiration, vegetation coefficient, soil moisture atfield capacity and at wilting point

Flow routing Slope, Manning’s roughness coefficient, hydraulic radius, averagetravel time and its standard deviation

Smedt et al. (2000) and Liu et al. (2003). For each grid cell, four layers are dividedas vegetation zone, root zone, transmission zone and saturated zone. The hydrolog-ical processes considered in the model are precipitation, interception, depression,surface runoff, infiltration, evapotranspiration, percolation, interflow, ground waterflow, and water balance in the root zone and the saturated zone (Table I). The totalwater balance for a raster cell is composed of the water balance for the vegetated,bare-soil, open water and impervious parts of each cell. This allows accounting forthe non-uniformity of the land use per cell, which is dependent on the resolution ofthe grid. The processes in each grid cell are set in a cascading way, which meansthat an order of occurrence of the processes is assumed after a precipitation event.The model predicts peak discharges and hydrographs, which can be defined forany numbers and locations in the channel network, and can simulate the spatialdistribution of catchment hydrological characteristics.

2.1. RUNOFF PRODUCTION

Three runoff components, surface runoff, interflow and groundwater flow, are con-sidered in the model. The model takes Hortonian flow as the main overland flowprocess, which occurs when rainfall intensity exceeds the infiltrability of soil. How-ever, for a complex terrain, particularly suburban and urban areas, precise estimationof infiltration parameters is rather difficult due to the high heterogeneity of the landand soil characteristics. Hence, simplified methods are still widely used by thehydrologists for surface runoff estimation in water resources planning, design andpractices, for instance the rational method and the soil conservation service (SCS)method.

In this study, a moisture-related runoff coefficient method is proposed for cal-culating surface runoff in each grid cell, which allows the actual runoff coefficientto vary in time, and in function of rainfall intensity, rainfall duration and cell


characteristics, giving an approximation to the surface runoff volume at each timestep. The initial losses due to interception and depression are considered separatelyin the formula

V = cscr(P − Ia) − Da (1)

where V [LT−1] is the surface runoff in depth over the time, P [LT−1] is the rainfallintensity, Ia [LT−1] is the interception loss, Da [LT−1] is the depression loss, cs [−]is a moisture related coefficient relying on the relative soil moisture content of theroot zone, and cr [−] is the potential runoff coefficient, which is assumed to dependupon slope, soil type, land use and the proportions of bare soil and impervious areasin a grid cell. The values of default runoff coefficients are taken from referencesfrom literature (Browne, 1990; Chow et al., 1988; Kirkby, 1978; Mallants andFeyen, 1990; Pilgrim and Cordery, 1993) and a table was generated, linking valuesof the runoff coefficient to slope, soil type and land use classes. The potentialrunoff coefficient is then the area-weighted average of the land use classes withinthe grid cell. The moisture related coefficient cs is time dependent and calculatedas a linear function of the soil moisture, in which the value is set to one when soilis saturated reflecting potential surface runoff conditions, and approaches to zerowhen soil moisture tends to zero. This is logical from a hydrological point of viewthat wet soil tends to produce more surface runoff, and dry soil tends to give moreinfiltration.

The first term on the right side of Equation (1) is the excess rainfall calculatedwith the net precipitation (P − Ia) multiplied by a potential runoff coefficientand a moisture related coefficient. The difference between net precipitation andexcess rainfall is the amount of infiltration into the soil. Surface runoff that isavailable for runoff routing is then computed with the excess rainfall subtracted bythe depression losses. cs is controlled not only by the antecedent soil moisture, butalso by the rainfall intensity and rainfall duration. Hence, high rainfall intensity orrainfall with long duration tends to give more percentage of runoff. The product ofcs and cr is the actual runoff coefficient, which varies both with time and rainfallintensity depending upon the soil moisture content, and allows computing excessrainfall for each time step during the model simulation.

The sum of interception loss and depression loss forms the initial abstraction,which does not contribute to runoff. In this study, interception loss is evaluated usinga simple reservoir model, in which the rainfall rate is reduced until the interceptionstorage capacity is satisfied. If the total rainfall during the first time increment isgreater than the interception storage capacity, the rainfall rate is reduced by thecapacity. Otherwise, all rainfall is intercepted in the canopy, and the remainder ofinterception is removed from the rainfall in the following time increments. Typi-cal interception capacity values are found in the literature, and a lookup table ofmaximum and minimum interception storage capacity corresponding to summerand winter periods was established linking values of interception storage capacity


to different land use classes. Thereafter, a simple sine-shaped variation curve wasproposed allowing the interception storage capacity to vary continuously with time.The intercepted water in the canopy is lost by evaporation and returns to the hydro-logical cycle with a potential evaporation rate.

Depression storage may have a considerable magnitude and plays an importantrole in flood modeling for small or medium storms. Due to the extreme variability ofthese characteristics, it is very difficult to specify a general relationship for the lossesdue to depression storage. In this study, the equation suggested by Linsley (1982)is used to calculate depression loss, in which the depression storage is assumedto be a function of depression storage capacity and increases exponentially withrainfall intensity up to the point where depression storage capacity is reached.This allows overland flow and depression storage to occur simultaneously. Thedepression storage capacity is a function of landform, soil type and vegetation.Based upon the typical values found in the literature, a lookup table of defaultdepression storage capacity is set up according to the category classes of slope,land use and soil type. However, local adjustment of these values is necessaryduring model calibration. Water held in depressions at the end of a rainfall eventdepletes either by evaporation with a potential evaporation rate or contributes tothe soil moisture.

2.2. WATER BALANCE

WetSpa has four water stores, i.e. the plant canopy, land surface, root zone and satu-rated zone, for each of which the water balance is maintained. The water balance inthe root zone is the most important one, as it controls the amount of surface runoff,interflow, evapotranspiration and groundwater recharge. Assuming that the ground-water table is below the root zone, the water balance in the root zone can be modeledfor each grid cell by equating inputs and outputs

D�θ

�t= P − I − V − E − R − F (2)

where D [L] is the root depth, �θ [L3L−3] is the change in soil moisture, �t [T] isthe time interval, I = Ia+Da , [LT−1], is the initial abstracts including interceptionand depression storage within time �t, E [LT−1] is the actual evapotranspirationfrom soil, R [LT−1] is the percolation out of the root zone, and F [LT−1] is theamount of interflow in depth over the time.

Evapotranspiration from soil and vegetation is calculated based on the rela-tionship developed by Thornthwaite and Mather (1955) as a function of poten-tial evapotranspiration, vegetation type, stage of growth and soil moisture con-tent. Potential evapotranspiration data can be obtained from field measurements,estimated from the historical records through statistical analysis, or calculatedwith Penman-Monteith equation when hourly meteorological data of net radia-tion, air temperature, relative humidity and wind speed are available. The process


of evapotranspiration is assumed to occur in order from interception storage, de-pression storage and soil subsequently. If the water content of interception storageand depression storage is equal to or greater than the potential evapotranspiration,all the evapotranspiration comes from those storages only. For the surface layer,actual evapotranspiration is computed as the area-weighted mean of the land useoccupation percentage, for which transpiration happens from the vegetated parts,evaporation happens from the bare soil, and there is no evaporation on imperviousareas. Finally, the total evapotranspiration is calculated as the sum of evaporationfrom interception storage, depression storage and the actual evapotranspiration fromthe soil.

Percolation and interflow are important components in the root zone water bal-ance. Both processes are assumed to be gravity driven. The rate of percolation orgroundwater recharge is determined by Darcy’s law in function of hydraulic con-ductivity and the gradient of hydraulic potential. When an assumption is made thatthe pressure potential only varies slightly in the soil, its gradient can be approxi-mated to zero, such that the percolation is controlled by gravity alone. Therefore,the percolation out of root zone is simply the hydraulic conductivity correspondingto the moisture content in the soil layer. The Brooks and Corey relationship betweenhydraulic conductivity and moisture content is used to define percolation, which issimply (Eagleson, 1978)

R = K (θ ) = Ks

(θ − θr

θs − θr

)(2+3B)/B

(3)

where K(θ ) [LT−1] is the unsaturated hydraulic conductivity, Ks [LT−1] is the sat-urated hydraulic conductivity, θ s [L3L−3] is the soil porosity, θ r [L3L−3] is theresidual moisture content, and B [−] is the soil pore size distribution index.

Interflow is assumed to occur in the root zone after percolation and becomessignificant only when the soil moisture is higher than field capacity. Darcy’s law andkinematic approximation are used to estimate the amount of interflow generatedfrom each cell, in function of hydraulic conductivity, the moisture content, the slopeangle, and the root depth

F = c f DS0 K (θ )/W (4)

where S0 [LL−1] is the surface slope, W [L] is the cell width, and c f [−] is a scalingparameter depending on land use, used to consider river density and the effectsof organic matter on the horizontal hydraulic conductivity in the top soil layer.Apparently, with Equation (4), rapid interflow will be generated in areas with highmoisture, steep slope and well vegetation. For other areas, little interflow will beproduced.

Since little is known about the hydro-geological conditions in the saturated layerat the cell level, the groundwater storage is simulated with a semi-distributed model


for simplicity, in which the groundwater balance is maintained in a small subcatch-ment scale with input of average groundwater recharge and output of groundwaterdischarge at the subcatchment outlet.

2.3. flOW ROUTING

The routing of overland flow and channel flow is implemented by the methodof the diffusive wave approximation. This method has been used in some recentGIS-based flood models (Fortin et al., 2000; Olivera and Maidment, 1999; Trochet al., 1994). A two-parameter response function, based on the average flow timeand the standard deviation of the flow time, is proposed in this study. The flowtime and its variance are determined by the local slope, surface roughness andthe hydraulic radius for each grid cell. The flow path response function at theoutlet of the catchment or any other downstream convergence point is calculated byconvoluting the responses of all cells located within the drainage area in the formof the probability density function (PDF) of the first passage time distribution. Thisrouting response serves as an instantaneous unit hydrograph and the total dischargeis obtained by a convolution integral of the flow response from all generated spatiallydistributed runoff.

Starting from the continuity equation and the St. Venant momentum equation,assuming one-dimensional unsteady flow, and neglecting the inertial terms and thelateral inflow to the flow element, the flow process can be modeled by the diffusivewave equation (Cunge et al., 1980)

∂ Q

∂t+ c

∂ Q

∂x− d

∂2 Q

∂t2= 0 (5)

where Q [L3T−1] is the discharge at time t and location x, t [T] is the time, x[L] is the distance along the flow direction, c [LT−1] is the location dependentkinematic wave celerity and is interpreted as the velocity by which a disturbancetravels along the flow path, and d [L2T−1] is the location dependent dispersioncoefficient. Assuming that the bottom slope remains constant and the hydraulicradius approaches the average flow depth for overland flow and watercourses, cand d can be estimated by c = (5/3)v, and d = (vH)/(2S0) (Henderson, 1966),where v [LT−1] is the flow velocity calculated with the Manning equation, and H[L] the hydraulic radius or average flow depth. Parameters c and d are assumedto be independent of the discharge, Q. Hence, the partial differential Equation (5)becomes parabolic, having only one dependent variable, Q(x, t). The hydraulicradius is determined by a power law relationship with an exceeding probability(Molnar and Ramirez, 1998), which relates hydraulic radius to the controlling areaand is seen as a representation of the average behavior of the cell and the channelgeometry

H = a(Ad)b (6)


where H [L] is the hydraulic radius, Ad [L2] is the drained area upstream of the cell,which can be easily determined by the flow accumulation routine in standard GIS,it a [-] is a network constant and b [-] a geometry scaling exponent both dependingon the discharge frequency. The parameters a and b can be estimated based on thewatershed characteristics.

An approximate solution to the diffusive wave equation in the form of a firstpassage time distribution was proposed by Liu et al. (2003), relating the dischargeat the end of a flow path to the available runoff at the start of the flow path

U (t) = 1

σ

√2π t3/t3

0

exp

[− (t − t0)2

2σ 2t/t0

](7)

where U(t) [T−1] is the flow path unit response function, t0 [T] is the flow time, andσ [T] is the standard deviation of the flow time. The parameters t0 and σ are spatiallydistributed, and can be obtained by integration along the topographic determinedflow paths as a function of flow celerity and dispersion coefficient

t0 =∫

c−1 dx (8)

σ =√∫

(2d/c3) dx (9)

The direct flow hydrographs at the basin outlet or any downstream convergentpoint are obtained by a convolution integral of the flow response from all contribut-ing cells

Q(t) =∫

A

∫ t

0V (τ )U (t − τ )dτd A (10)

where Q(t) [L3T−1] is the direct flow hydrograph, V(τ ) [LT−1] is the surface runoffinput in a grid cell, τ [T] is the time delay, and A [L2] is the drainage area of thewatershed.

In this study, interflow is assumed to contribute to the surface runoff at theoutlet of each cell, and routed to the catchment outlet together with surface runoffwithout redistribution among downslope cells for simplicity. Groundwater flowis modeled with a linear reservoir method on small subcatchment scale, while anon-linear reservoir method is optional in the model with storage exponent of 2(Wittenberg and Sivapalan, 1999). The groundwater outflow is added to any runoffgenerated at the subcatchment outlet to produce the total streamflow. Hence, theflow routing consists of tracking runoff along its topographic determined flowpath, and evaluating groundwater flow for each small subcatchment. In order toconsider the damping effect of the river, overland flow and interflow are routed


firstly from each grid cell to the main channel, and joined with groundwater flow atthe subcatchment outlet. Then the total hydrograph is routed to the basin outlet bythe channel response function. The total discharge is the sum of the overland flow,interflow and groundwater flow.

3. GIS Implementation

WetSpa is tightly coupled with GIS as a raster model. Input base maps include DEM,soil type and land use map. Besides, the digital information of stream and rainfallgauging sites, official watershed boundary and river network, sewer systems, mainhydraulic and civil infrastructures are necessary geo-referenced data for a complexterrain. Once the project database is setup, the work begins by analyzing data,extracting information, producing parameter maps, and running the distributedmodel. GIS functions greatly enhance the capabilities for watershed descriptionand interpretation by means of powerful distributed indicators, which accounts notonly for the watershed characteristics but also for their distribution and individuallocalization in space.

3.1. DRAINAGE SYSTEM

Elevation data in the form of a DEM are the principal digital data source foracquiring watershed properties in the GIS-based WetSpa model. The raster-typeDEM has to be compatible with remotely sensed data layers such as land use andsoil type. Based on the digital elevation model, hydrological GIS tools are used toextract information on the watershed boundaries, such as slope, flow length, flowdirection and accumulation, configuration of stream network, subwatershed, etc.,providing a suitable framework for the modeling approach. The flow accumulationmap is used for synthetic network extraction, where each pixel is associated toan upstream drainage area. The distinction between hill slopes and channel pathscan be achieved simply by fixing a threshold drainage area for which the flowconcentration is sufficient to initiate a channel.

In a natural drainage basin with very little or no human interactions, aspect infor-mation obtained from a DEM alone is a good indication of flow direction, and thederived internal drainage structure of a watershed can be a perfect reflection of thenatural reality. However, in complex terrain, such as urban or suburban watersheds,sewer systems, roads, artificial channels, etc., are important elements in drainagestructure configuration, and often govern the flow direction more strongly than thederived slopes at local scale. Since most of these features are not sufficiently repre-sented in a DEM, additional procedures for deriving more realistic flow directionmap have to be performed using GIS overlaying techniques. A general flow direc-tion map can be generated using geographic data alone. Thereafter, direction mapsof sewer areas, main water routes and a fine river network can be created separatelybased on the DEM and available coverage maps. Finally, the general flow directionmap can be overlaid by the flow direction map of sewers, drain ditches, and streams


subsequently. In such a way, normal flow directions are altered fundamentally bythe presence of artificial drainage systems. The derived flow direction map is thenused for further drainage structure delineation.

3.2. SOIL AND LAND USE

In addition to the topography, soil and land use properties are utilized in the WetSpamodel to specify the land surface characteristics that determine the partitioning ofincident rainfall into infiltration and runoff, as well as the simulation of subsurfaceflow and the vertical water and energy budget. Parameters, which depend upon soiltype and land use, are incorporated in the model as attribute tables of the land use andsoil type maps. The soil texture is classified based on the US Department of Agricul-ture classification in the model for the identification of soil dependent parameters,such as porosity, field capacity, permanent wilting point, residual moisture content,pore size distribution index, saturated hydraulic conductivity, etc. An assumptionmade in the model is that soil hydraulic properties remain constant throughout theroot zone for each grid cell. Default values are available in the model, but can beadjusted by the user to more appropriate values in case more specific informationis available.

The digital land use map is normally obtained from a high-resolution remotelysensed image. Fourteen land use classes are identified in the WetSpa model, whichare significantly different from each other on the basis of their effects on hydrologi-cal processes. Each of these classes is characterized by quantitative attributes, as forinstance, canopy resistance, albedo, root depth, interception capacity, Manning’sroughness coefficient, etc. The proportion of impervious area and bare soil areconsidered separately for setting up the overall land use dependent parameters foreach grid cell. Using the roughness map derived from the land use, combined withthe slope map and hydraulic radius map derived from the DEM, the average flowvelocity, celerity and dispersion coefficient can be calculated for each cell. Con-sequently, the average travel time and standard deviation are obtained using theweighted GIS flow length routine, and used for determining flow response functionfor each grid cell. Afterwards, the land use is regrouped into six classes, namelyforest, pasture, crop, bare soil, urban and open water surface, for deriving potentialrunoff coefficient and depression storage capacity in combination with slope andsoil texture classes.

3.3. SPATIAL INPUT AND OUTPUT

Rainfall and potential evapotranspiration are the two meteorological variablesneeded for the WetSpa model. If the Penman-Monteith equation is used for estimat-ing potential evapotranspiration, the data of net radiation, air temperature, relativehumidity, and wind speed are required. The method of Thiessen polygon is appliedto interpolate precipitation and other meteorological variables observed at differ-ent meteorological stations. This implies that the rainfall or other meteorological


variables of each grid cell is set to the value recorded at the nearest gauging site.The average rainfall and potential evapotranspiration for each subwatershed can beestimated by integration of the values on the cells belonging to that subwatershed.

The WetSpa model computes and generates time series of flow hydrographs atcertain points and maps of spatial outputs. These maps include interception, surfacerunoff, infiltration, soil moisture, actual evapotranspiration, percolation, interflow,as well as distributed gross and net rainfall intensity in function of time. For each ofthese, maps in GIS format can be saved at a specified time increment, which can beused for graphical presentation to see the complete temporal and spatial variationof each of the above state variables during a model simulation.

4. Model Application

4.1. WATERSHED DESCRIPTION AND DATA AVAILABILITY

The model is tested on the 67.8 km2 Barebeek catchment, which is a small watershedof the Dijle river basin, Belgium. It is a typical suburban area situated northeastof Brussels. The Brussels international airport is located in the upper area of thecatchment. Four main traffic lines cross the watershed in different directions, andmany country roads crisscross the area from one village to another. The Leuvencanal passes through the area in the north. A small lake covering about 0.55 km2 islocated near the basin outlet. Several residential areas with sewer systems exist in thewatershed occupying about 28% of the total area. The watershed drainage system,together with the main civil infrastructures and measuring stations are presented inFigure 1(a).

Figure 1. (a) Gauging network and drainage system of the study area, and (b) Land use mapof the study area.


The study area is discretized into grid cells of 50 m by 50 m. The topography isdigitalized from 1/10,000 maps and the soil types are obtained from the physicalsystem map of Flanders. The catchment is rather low and flat with average basinslope of 0.63%. Elevation differences are small with extreme values ranging from 5to 68 m. The dominant soil types are sandy loam (66.2%) and loamy sand (29.8%),while the rest are sand, loam, silt loam and clay scatted around the catchment. Thesoil cover is obtained from the digital land use map of Flanders, which is basedon remote sensed data of 1995. The resampled land use map for use in the modelis presented in Figure 1(b). The study area is well vegetated. Forrest (16.8%) ispredominant in the river valleys, while the higher terrains consist of agriculturalareas, with pasture (24.7%) and crops (36.9%), strongly intermixed with urbanareas (16.2%), as villages, roads and Brussels international airport in the south.

The study area has a maritime temperate climate with no proper distinction be-tween rainy and dry seasons. Rainfall is relatively uniformly distributed throughoutthe year, but storms have low intensities and long durations in winter, and are in-tense and of short duration in summer. High runoff occurs in winter and low runoffin summer due to dry soils and high evapotranspiration. Floods happen frequentlyin winter with storms strongly influenced by the westerly atmospheric fluxes thatbring humid air masses from the Atlantic Ocean. Heavy storms usually last 2–3days with peak rainfalls concentrated into 3–6 h. The average annual precipitationin the region is about 800 mm, and the annual potential evaporation from free watersurface is around 650 mm.

During the period of December 1998–February 1999, an intensive hydrologicalresearch was carried out to study the water quantity and quality in the Barebeekwatershed. Two temporary rainfall gauges, P1 and P2, and five stream gauges,MO1–MO5, were set up as shown in Figure 1(a). MO6 is a regular flow moni-toring station located in the downstream part of the river with a relatively longdischarge record. The main meteorological station is situated at Ukkel, locatedsouth of Brussels about 12 km away from the catchment. At this station, a verylong meteorological record starting from the year 1898 is available, which can beused for model simulation in the study area.

4.2. MODEL PARAMETER OPTIMIZATION

Calibration of the model was performed using the measured precipitation and dis-charge data during the period of December 1998–February 1999. The potentialevapotranspiration data was obtained from Ukkel station. Predicted hydrographswere compared with the measured streamflows at each stream gauge. Model cali-bration was performed at the most downgauging site MO6, while simulation resultsat other stream gauging sites can be seen as model verifications. Further verificationwas carried out at MO6 for the period of September 1998–December 1999, consid-ering the seasonal parameter verification and using the available precipitation andpotential evapotranspiration data of Ukkel.


Figure 2. (a) Distribution of potential runoff coefficient, and (b) Distribution of average flowtime to the catchment outlet.

Figure 2(a) shows the distribution of potential runoff coefficients that result fromthe different slope, soil type and land use class combinations. The watershed is cov-ered by either dense plants or impervious areas. Due to the model grid size, cellsmay not be 100% impervious in reality. In this study, the percentage of imperviousarea in a grid cell is computed based on land use classes, with 30% for residen-tial areas, 70% for commercial and industrial areas and 100% for streams, lakesand roads. Default runoff coefficients for these areas are calculated by adding theimpervious percentage with a grass runoff coefficient multiplied by the remainingpercentage. This results in potential runoff coefficients of 40–100% in urban areas,while other areas have much smaller values, down to 3% for forests in valleys withpractically zero slopes.

A flow direction map is generated considering the effect of sewer systems. Theaverage hydraulic radius is computed for normal floods, i.e. with a return period of2 years. The parameter a and b are set to 0.07 and 0.43 respectively resulting in aminimum hydraulic radius for overland flow of 0.005 m and a maximum hydraulicradius for channel flow of 0.5 m at the watershed outlet. The value of parameters aand b can be increased for more extreme floods. Next, the overland flow velocity iscalculated with Manning’s equation. The urban areas have a remarkable influence,due to the artificial drainage facilities as sewer systems that result in flow velocitiesof order 0.5 m/s or more. On the other hand, the overland flow velocity in valleyareas is very small, due to the high resistance of the soil cover, being mostly forest,and the very faint slopes. With this information the flow path response functions arecalculated for each cell using the diffusive wave approximation. Figure 2(b) showsthe resulting average flow time from each grid cell to the outlet of the watershed,in which the average flow time is less than 10 h for the main river and up to 40 hfor the most remote areas.


Figure 3. (a) Observed vs. calculated flow hydrographs at MO6, and (b) Observed vs. calculatedflow hydrographs at MO3.

A comparison of calculated and observed discharges at the most downstreammeasuring station MO6 is presented in Figure 3(a). A comparison at an internalsite MO3 for the same period is shown in Figure 3(b), while results for the othermeasuring stations are similar. Six storms occurred during this calibration period,with quite small intensities lower than 3 mm/h, but lasting for relatively long periods,i.e. 2–5 days for each. Flow volumes are mainly determined by the baseflow, about82% for MO6 and 71% for MO3. Peak discharges result from surface runoff duringeach flood, and are mainly generated in urban areas.

A comparison at MO6 for the period of September 1998–December 1999 ispresented in Figure 4. One can notice a reasonable agreement between the sim-ulation results and the observed hydrograph. Peaks in the hydrograph are ratherwell simulated, as well as for their shape and time of occurrence. From the 16months simulation results, about 11% of the total rainfall is lost by interception,53% of the total rainfall returns to the atmosphere as evapotranspiration, and 32%

Figure 4. Observed vs. calculated flow hydrographs at MO6 for the period of September1998–December 1999.


is recharged to the groundwater, which mainly happens during the winter season.The simulated flow volume is composed of surface runoff (28%), groundwater flow(70%), and interflow (2%). Interflow is not an important flow component in thisstudy area, due to the fact that the land slope is too small for the occurrence oflateral flow in the unsaturated zone. Infiltrated water either stays in the soil and islost as evapotranspiration, or is recharged to the saturation zone for generation ofbaseflow. Floods occur frequently in the winter season, because of saturated soilsand high baseflow, even though the storms were not very intensive. The largeststorm occurred on August 10, 1999, with maximum rainfall intensity of 15 mm/h,and the accumulative rainfall was 39 mm within 15 h. This storm did not cause asevere flood, since the antecedent soil moisture of the watershed was much lowerthan the field capacity, thus leading to very high infiltration and storage in the soil.Additionally, a large amount of rainfall was lost by interception and depressionstorage during the initial phase. The measured peak discharge was 1.43 m3/s, andthe calculated peak discharge was 1.82 m3/s.

Figure 5(a) shows the simulated spatial distribution of relative soil wetness (θ /θ s)on October 8, 1999, before the main storm. It is found that higher soil wetness waspresent in river valleys and lower wetness in the upper areas. In general, the soilmoisture content is quite small in the watershed due to little precipitation and highevapotranspiration in the previous month. Additionally, it is seen from the figurethat the relative soil wetness in the downstream areas close to the river outlet wasmuch lower than that in other areas. This is because the area is covered by sandysoils with rather small water holding capacity, and most soil water contributed to theevapotranspiration and percolation. Figure 5(b) gives the simulated surface runoffproduced in the following hour. High surface runoff was mainly generated from

Figure 5. (a) Simulated relative soil saturation on 8/10/1999 4:00, and (b) Simulated surfacerunoff on 8/10/1999 4:00–5:00.


the impervious areas and open water surfaces, while surface runoff for other areaswas rather small, especially for the areas covered by forest and sandy soils.

The model performed well with the Nash-Sutcliffe efficiency criteria for thecalibration period, with the total water balance 2% under estimated, the efficiencyfor reproducing river discharges 72.4%, the ability to reproduce low flow and highflow 83.8% and 76.8% respectively. This indicates that the model is suitable forboth peak flow prediction and hydrograph simulation in this watershed.

4.3. MODEL APPLICATION USING THE HISTORICAL AND IDF DATA

To demonstrate the usefulness and performance of the model, a historical 102-yearseries of precipitation data from Ukkel was processed under the present land usecondition. The daily and hourly potential evapotranspiration was estimated fromthe historical records of Ukkel through statistical analysis. The resulting hydro-graphs for the whole catchment were then analyzed statistically to determine thecharacteristics of peak discharges.

Also, hourly design storms developed by Willems (2000) were introduced asrainfall input to the model for calculating corresponding design floods to comparewith the modeling results of the 102-year rainfall series. The IDF relationships wereestablished based on the long rain gauge record of 10-min precipitation depths forthe period 1967–1993 at Ukkel, which is the same rainfall station for the modelsimulation of the Barebeek. Storms of two different types, air mass thunderstormsand cyclonic/frontal storms, are separated based on their distribution of peak-over-threshold intensity. This is done for each duration in the range 10 min 15 days,using a two-component exponential distribution. Different mixtures of the two typestorms are estimated for summer and winter conditions.

The hourly design summer and winter storms with 15 days duration and 100-year return period are shown in Figure 6(a), in which the peak storm intensities

Figure 6. (a) Design storms with 100-year return period, and (b) Simulated floods with 100-year design storms.


are 36.3 mm/h and 13.1 mm/h respectively for summer and winter design storms,while the storm volumes are the same, i.e. 174 mm. Summer design storms are moreintensive corresponding mostly to air mass thunderstorms, and winter design stormsare less intensive corresponding mostly to cyclonic/frontal storms. The resultingflood hydrographs from the 100-year design storms were simulated with the WetSpamodel, as shown in Figure 6(b), resulting in peak discharges of 5.96 m3/s and8.55 m3/s respectively for the design summer and winter storms. The winter floodswere simulated with the assumption that the initial soil moisture content was equalto the field capacity, while for summer design floods, the initial soil moisture contentwas assumed to be half of the field capacity. Figure 6(b) shows clearly that the designsummer storm produces higher peak discharge and lower flood volume comparedwith that induced from the design winter storm due to the high rainfall intensity,high evapotranspiration rate, and low antecedent soil moisture. These results reflectthe typical pattern of summer and winter floods of the catchment on one hand andthe importance of soil moisture in controlling the runoff production on the otherhand.

The comparison of design peak flows for the two methods is presented in Fig-ure 7(a). For each return period, the design winter peak flow is smaller than thedesign summer peak flow, but is getting closer with each other for short returnperiods. The statistical result of the 102-year series model simulation is very closeto the result calculated from the summer design storms. This is because the annualmaximum flood occurs mostly in the summer season.

The simulated largest flood at the watershed outlet on August 30, 1996 is pre-sented in Figure 7(b). It shows a typical pattern that is present in most of theprecipitation events in the study area that lead to flood discharges. The actual stormwas preceded by another storm 5 days before the main event. The simulated an-tecedent soil moisture was very low at that moment with average moisture contentaround 43% of the saturation capacity. This resulted in a very small actual runoffcoefficient for pervious soils, and the surface runoff was mainly produced from the

Figure 7. (a) Comparison of the design peak flows, and (b) Simulated maximum flood at thecatchment outlet.


impervious areas. This storm did not lead to flooding, but increased the root zoneaverage soil moisture to field capacity, or about 54% of the saturation capacity. OnAugust 29, 1996, a storm event with 108 mm fell on the entire catchment, causinga severe flood in the watershed. The calculated surface runoff was generated fromevery grid cell of the catchment. Complete saturation of the root zone occurred inthe river valleys and the areas with sandy soils, forest cover and very flat slopes, butnot on the entire watershed. The average moisture content of the watershed was 70%of the saturation capacity after the storm event. This information, together with thesimulated stream flow hydrographs, gives a more complete view of the hydrologicalbehavior and allows a better understanding of the hydrological processes.

5. Summary and Conclusions

A physically based distributed hydrological model compatible with remote sensingand GIS is presented for simulating the hydrological behavior of a watershed. Thegeneration of runoff depends upon rainfall intensity and soil moisture status and iscalculated in function of slope, land use and soil type. The runoff is subsequentlyrouted through the watershed along flow paths determined by the topography us-ing a diffusive wave transfer model that leads to response functions between anystart and end point, depending upon slope, flow velocity and dissipation character-istics along the flow paths. The model can predict not only the flood hydrographat any controlling point of the river, but also the spatially distributed hydrologicalprocesses, such as surface runoff, soil moisture, interflow, groundwater recharge,actual evapotranspiration and so on, at each time step during a simulation.

The modified runoff coefficient for surface runoff production is no longer theconventional runoff coefficient used in the rational method, but a measure of rainfallpartitioning capacity, which allows varying with time, rainfall intensity, rainfallduration and the cell geophysical characteristics. With such improvements, themodel can be used for the computation of storm hydrographs for any size watershed.Processes in water and energy transfer between soil, plant and atmosphere aresimulated using simplified equations for each grid cell. Among these, water balancein the root zone is important, as the moisture in the root zone is a key factorto control the amount of surface runoff, interflow, actual evapotranspiration andgroundwater recharge. All model parameters can be obtained from DEM, land useand soil type data of the watershed or combinations of these three fundamentalmaps. The spatial distribution of rainfall intensity, potential runoff coefficient andthe antecedent soil moisture content are governing factors of the flood volume,while the hydraulic radius and the channel roughness coefficient are sensitive forflow routing simulations.

GIS provides a powerful platform for developing the model, calibrating parame-ters and displaying model results in a spatial way, so that it is possible to capture localcomplexities of a watershed and compare model results to the field measurements.The model was validated on a small watershed in Belgium for which topography


and soil data are available in GIS form, while the land use and soil cover was ob-tained from remote sensed images. The resulting calculated hydrographs comparefavorably with measurements. The usefulness and utility of the model are subse-quently demonstrated by forecasting peak discharges resulting from an observed102 years precipitation series. The resulting discharges were analyzed statisticallyto determine the characteristics of extreme flood events and compared with theresults computed from design storms. Comparison of the two methods shows thatthe model is capable to predict both normal and extreme floods.

As discussed in the paper, the model makes full use of the remote sensed dataand calculations are for the most part performed by standard GIS tools, such that themodel is especially useful for flood prediction on complex terrain and analyzing theeffects of topography, soil type, and land use or soil cover on the flood. Additionally,the model can be easily coupled with other water quality and soil erosion models,and used for simulating spatial hydrological behavior of a river basin.

References

Abbott, M. B., Bathurst, J. C., Cunge, J. A., O’Connell, P. E., and Rasmussen, J., 1986, ‘An introductionto the European hydrological system – Système hydrologique européen (SHE), Part 2: Structureof a physically-based, distributed modeling system’, J. Hydrol. 87, 61–77.

Beven, K. J. and Kirkby, M. J., 1979, ‘A physically based variable contributing area model of basinhydrology’, Hydrol. Sci. Bull. 24(1), 43–69.

Browne, F. X., 1990, ‘Stormwater Management’, in R. A. Corbitt (ed), Standard Handbook of Envi-ronmental Engineering, McGraw-Hill, New York, pp. 7.1–7.135.

Chow, V. T., Maidment, D. R., and Mays, L. W., 1988, Applied Hydrology, McGraw-Hill, New York.Cunge, J. A., Holly, F. M., and Verwey, A., 1980, Practical Aspects of Computational River Hydraulics,

Pitman, London, p. 45.De Smedt, F., Liu, Y. B., and Gebremeskel, S., 2000, ‘Hydrological modeling on a catchment scale

using GIS and remote sensed land use information’, in C. A. Brebbia (ed), WTI press, Boston,pp. 295–304.

Downer, C. W., Ogden, F. L., Martin, W., and Harmon, R. S., 2002, ‘Theory, development, andapplicability of the surface water hydrological model CASC2D’, Hydrol. Process. 16, 255–275.

Eagleson, P. S., 1978, ‘Climate, soil, and vegetation, a simplified model of soil moisture movementin liquid phase’, Water Resour. Res. 14(5), 722–730.

Ewen, J., Parkin, G., and O’Connell, P. E., 2000, ‘SHETRAN: Distributed river basin flow andtransport modeling system’, J. Hydrol. Eng. 5, 250–258.

Fortin, J. P., Turcotte, R., Massicotte, S., Moussa, R., Fitzback, J., and Villeneuve, J. P., 2001, ‘Adistributed watershed model compatible with remote sensing and GIS data, I: Description of themodel’, J. Hydrol. Eng. 6, 91–99.

Henderson, F. M., 1966, Open Channel Flow, McMillan, New York, p. 522.Kirkby, M. J. (ed.), 1978, Hill-Slope Hydrology, John Wiley & Sons, Chichester, New York, USA, p.

235.Linsley, R. K. Jr., Kohler, M. A., and Joseph Paulhus, L. H., 1982, Hydrology for Engineers, McGraw-

Hill, New York, p. 237.Liu, Y. B., Gebremeskel, S., De Smedt, F., Hoffmann, L., and Pfister, L., 2003, ‘A diffusive transport

approach for flow routing in GIS-based flood modeling’, J. Hydrol. 283, 91–106.Mallants, D. and Feyen, J., 1990, ‘Kwantitatieve en kwalitatieve aspecten van oppervlakte en grond-

waterstroming (in Dutch)’, Vol. 2., KUL, p. 96.


Molnar, P. and Ramirez, J. A., 1998, ‘Energy dissipation theories and optimal channel characteristicsof river networks’, Water Resour. Res. 34, 1809–1818.

Olivera, F. and Maidment, D., 1999, ‘Geographic information systems (GIS)-based spatially dis-tributed model for runoff routing’, Water Resour. Res. 35, 1155–1164.

Pilgrim, D. H. and Cordery, I., 1993, ‘Flood runoff’, in D. R. Maidment (ed), Handbook of Hydrology,McGraw-Hill, New York, pp. 9.1–9.42.

Quinn, P., Beven, K., Chevallier, P., and Planchon, O., 1991, ‘The prediction of hillslope flow pathsfor distributed hydrological modelling using digital terrain models’, Hydrol. Process. 5, 59–79.

Troch, P. A., Smith, J. A., Wood, E. F., and De Troch, F. P., 1994, ‘Hydrological controls of largefloods in a small basin’, J. Hydrol. 156, 285–309.

Wang, Z., Batelaan, O., and De Smedt, F., 1997, ‘A distributed model for water and energy transferbetween soil, plants and atmosphere (WetSpa)’, Phys. Chem. Earth 21, 189–193.

Willems, P., 2000, ‘Compound intensity/duration/frequency-relationships of extreme precipitation fortwo seasons and two storm types’, J. Hydrol. 233, 189–205.

Wittenberg, H. and Sivapalan, M., 1999, ‘Watershed groundwater balance estimation using streamflowrecession analysis and baseflow separation’, J. Hydrol. 219, 20–33.

flood modeling for complex terrain using gis

Documents

hydrological processes

hydrological modeling

distributed models

hydrological behaviorof

land use data

soil data

hydrological cycle

basin characteristics