Scales in Hydrology and Water Management I Echelles en hydrologie et gestion de l'eau
(IAHS Publ. 287. 2004)
97 Prediction of statistical scaling in peak flows for rainfall-runoff events: a new framework for testing physical hypotheses VI.IAY K. GUPTA Department of Civil and Environmental Engineering, Cooperative Institute for Research in Environmental Sciences, University of Colorado, Boulder, Colorado 80309, USA gupt av(S>. ci res, col orado. edu Abstract Key elements of a new physical theory are explained which predicts statistical scaling in floods on nested sub-basins for individual rainfall-runoff events. The theory combines physical processes with the geometty and topology of channel network patterns and statistical variability. A discrepancy between model predictions and empirically observed scaling relationships provides a basis for testing physical assumptions without calibrating model parameters. An idealized example is used to illustrate how model predictions can be tested against empirical scaling relationships using topographic and flood data sets for individual rainfall-runoff events from the Goodwin Creek basin, Mississippi, USA. Three key problems are addressed here: (a) scale, (b) linking observed space-time statistical variability with underlying physical processes involving a large number of dynamic parameters, (c) the existence of major gaps between data needs and data availability for testing the scaling theory. Research on scaling in floods was instigated by the classic unsolved problem of prediction in ungauged and poorly gauged basins. Ke y wo r d s event - bas ed flood pr edi ct i on; spat i al var i abi l i t y; st at i st i cal s cal i ng; unga uge d bas i ns Prvision de la scalance statistique des dbits de pointe des vnements pluie-dbit: un nouveau cadre pour les tests d'hypothses Rsum On prsente ici les lments fondamentaux d'une nouvelle thorie physique prvoyant la scalance statistique des crues pour des vnements pluie-dbit individuels dans des bassins embots. La thorie combine les processus physiques avec la gomtrie et la topologie des types de rseau de chenaux d'coulement et leur variabilit statistique. Une diffrence entre les prvisions du modle et la scalance observe empiriquement fournit une base pour tester les hypothses physiques utilises sans avoir caler les paramfres. Nous nous sommes bass sur un exemple schmatique pour montrer comment les prvisions du modle peuvent tre testes vis vis des relations de scalance empiriques en utilisant les donnes topographiques et de dbits d'vnements pluie-dbit individuels du bassin de la Goodwin Creek (Mississipi, USA). Trois problmes fondamentaux sont abords ici: (a) les chelles, (b) la liaison entre la variabilit statistique spatio-temporelle observe et les processus physiques sous-jacents qui dpendent d'un grand nombre de paramtres dynamiques, (c) l'existence d'un important hiatus entre le besoin et la disponibilit des donnes ncessaires pour tester la thorie de l'invariance d'chelle. Cette recherche sur la scalance des crues a t inspire par le problme classique encore irrsolu de la prvision dans les bassins non ou peu jaugs. Mo t s cl efs s cal ance s t at i s t i que; pr vi s i on d' un v ne me nt de cr ue; var i abi l i t s pat i al e; bas s i ns non j a ug s 98 Vijay K. Gupta INTRODUCTION Many practical problems need prediction of peak flows or floods at a wide range of space and time scales in drainage basins. Practical interest in floods comes from a variety of sources in addition to engineering design. For example, the variation of flows in terms of their magnitude and frequency, or regime, is a primary factor that controls channel form and process and the nature of aquatic and riparian ecosystems. It causes stream channels and ecosystems to co-exist in a constant state of flux. More- over, the formation of a channel network and its ecosystems are a direct result of the culminated history of flow events and their future condition depends upon the characteristic flow regime of a basin (Poff et al, 1997; Beschta et al., 2000; Wolman & Miller, 1960). This paper explains some basic elements of a physical theory of scaling in spatial flood statistics on nested channel networks that has been developing since Gupta et al. (1996) published the first paper about it. Results in Gupta et al. (1996) have been generalized in many directions on idealized mean self-similar channel networks (Menabde et al., 2001; Menabde & Sivapalan, 2001 ; Troutman & Over, 2001). However, there is a pressing need to further generalize and test the scaling theory of floods on real channel networks. Without this key step, it will not be possible to apply this new theory to real world situations and substantially advance the existing technology for flood prediction in poorly gauged and ungauged basins. Such a generalization will need to incorporate the presence of natural spatial variability in the topological, hydraulic and the hydrological properties of real channel networks, and space-time variability of rainfall, soil moisture and runoff generation. Moreover, nested networks with a large number of streamgauges are required for testing the predictions of the scaling theory. Unfortunately, such basins are quite limited in the USA and the rest of the world. These issues make the problem of generalization of the scaling theory from idealized to real networks a major scientific challenge. Fortunately, a large number of streamgauges and raingauges are available on two experimental basins of the US Department of Agriculture (UDSA), the Walnut Gulch basin (Arizona) and the Goodwin Creek basin (Mississippi). We are using these two basins to develop and test the scaling theory of floods on real networks (Mantilla et al., 2004; Fury & Gupta, 2004). In this paper, I use the topographic and the stream flow data from the Goodwin Creek basin to illustrate some key elements of the scaling theory of floods for individual rainfall-runoff events. In its simplest form, statistical simple scaling connects two (joint) probability distributions at any two arbitrary scales by a power law. For simplicity, let us take drainage area as a spatial scale parameter. Equality in distributions implies that the mean and higher-order finite moments of flood peaks, or the quantiles, can be represented as power laws with respect to drainage area (Gupta & Waymire, 1998a,b). Power laws describe similarity or similitude. For example, power laws have served as the foundation for dimensional analysis and dynamic similitude in fluid mechanics (Barenblatt, 1996, Chap. 1). Contemporary scientific literature dealing with geometrical self-similarity and fractals is based in power laws (Barenblatt, 1996; Mandelbrot, 1982). The concept of self-similarity and its generalizations are having a great impact on hydrology (Sposito, 1998), geosciences (Turcotte, 1997), and other sciences such as fluid mechanics (Barenblatt, 1996). Prediction of statistical scaling in peak flows for rainfall-runoff events 99 Three types of empirical statistical scaling analyses for floods have been published. The first considers scaling in observed peak flows for individual rainfall- runoff events on nested channel networks (Ogden & Dawdy, 2003). The second considers scaling in annual peak flow statistics on nested channel networks (Goodrich et al., 1997), and the third considers scaling in annual peak flow statistics in un-nested basins within a homogeneous region (Cathcart, 2001; Gupta & Dawdy, 1995; Smith, 1992). Analyses of peak flows for individual rainfall-runoff events on nested channel networks provide a natural starting point for developing a physical understanding of statistical scaling. In this paper, I have chosen to illustrate how channel network geometry involving scaling in width function maxima enters into the scaling theory of floods for individual rainfall-runoff events. Scaling in width function maxima is illustrated here for the Goodwin Creek basin. It is empirically observed for many real networks that have been analysed by Veitzer & Gupta (2001), who also found scaling in the maxima of width functions for a new class of channel network models called Random Self-Similar Networks (RSN) (Veitzer & Gupta, 2000). As illustrated here, scaling in width function maxima provides a natural starting point for understanding how space-time variable physical processes transform rainfall to floods that exhibit statistical scaling on real networks. The scaling theory of floods predicts empirically observed statistical scaling parameters under a set of assumptions about physical processes and channel network geometry. Prediction allows us to test different physical assumptions and hypotheses within a rigorous mathematical framework without calibrating and fitting model parameters. This is the main conceptual issue that I wish to illustrate in this paper through a simple example. Klemes (1997) has thoughtfully discussed the issue of testing a model using data for advancing physical understanding versus curve fitting a model to data. Testability in scaling theory serves as a major point of departure of this approach from a large number of hydrological models that are fitted to data; see Singh (1995) for many examples of such hydrological models. Research on the scaling theory of floods was instigated by the long-standing need to develop a fundamental physical basis for purely statistical approaches to regionalization of floods in ungauged and poorly gauged basins (Cathcart, 2001; Gupta et al., 1994; Smith, 1992). Almost all the basins in the world are either ungauged or poorly gauged. A major long-term objective of this theory is to provide a process-based predictive understanding of flood scaling statistics in nested and un-nested ungauged basins ranging from the time scales of individual events to seasonal, annual, inter-annual and longer time scales. Gupta & Waymire (1998a) provided a self-contained exposition of the developments until 1996, and Gupta (2004) published a brief review of further progress until recently. In recognition of the global importance of the basic problem of prediction in ungauged basins (PUB) to hydrological sciences and engineering, and to closely related sciences, IAHS has launched a new decadal initiative on PUB (Sivapalan et al., 2003). The scaling theory of floods addresses three key problems in river basin hydrology that have been articulated to varying degrees in the literature: (a) the problem of scale; (b) the problem of linking observed space-time statistical variability with underlying physical processes at multiple scales. Following Gupta (2004), I discuss here how the scaling theory tackles the problem of a very large number of dynamic parameters that are needed to model spatially variable physical processes in rainfall-runoff relationships; 100 Vijay K. Gupta and (c) the problem that major gaps exist between data needs and data availability for developing and testing the scaling theory. The significance of these three problems individually and collectively is not limited to floods but applies to many hydrological variables, e.g. rainfall and vapotranspiration. Collectively, these three problems can be regarded as a central challenge for hydrology. I will focus on explaining these key issues through a simple example from the scaling theory of floods, and refer to the published literature for many technical details of the theory (Ogden & Dawdy, 2003; Menabde et al, 2001; Menabde & Sivapalan, 2001; Troutman & Over, 2001; Veitzer & Gupta, 2001; Reggiani et al, 2001). This paper has the dual objective of providing a clear road map for understanding a newly growing body of literature on the scaling theory of floods (Gupta, 2004). This paper is organized as follows. The next section begins with a spatially discrete equation of mass balance at the hillslope-link scale in a basin, and the scale issues that arise in representing physical processes in this equation. This is followed by a brief discussion of the parametric complexity that ensues due to spatial variability of the physical processes that transform rainfall to runoff. The next section explains some elements of the scaling theory through an idealized example and illustrates how model predictions can be tested against data. Our approach is based on predicting the empir- ically observed spatial statistical scaling in peak flows for rainfall-runoff events. Deviations between model predictions and empirical scaling relationships provide a new framework for testing physical assumptions and hypotheses. However, computation of empirical scaling relationships requires stream flow data at several locations in a nested basin, which are rarely available in the USA or globally. I briefly discuss our approach for solving this problem in a separate section. Scaling provides a spatially distributed metric, based in similarity, for testing model predictions of peak flows. It should be contrasted with the minimum mean square error between predicted and observed flow hydrographs at a basin outlet, which represents a spatially inte- grated metric that is widely used in calibrating rainfall-runoff models. I close with a brief discussion of future research. SCALE PROBLEMS IN FORMULATING A MATHEMATICAL THEORY OF FLOODS The mathematical formulation begins with a discrete mass-balance or continuity equation and uses the geomorphic decomposition of a drainage basin as a collection of channel links and hillslopes as shown in Fig. 1. Shreve (1966) introduced the concept of a link in his well-known theory of channel networks known as the random model. Each link of a river network is surrounded by two hillslopes, one on each side of it, which are shown in the same colour in Fig. 1. Precipitation on each hillslope produces runoff into its adjoining link, and it involves the physical processes of infiltration, vapotranspiration, and overland and subsurface flows. A channel network covers spatial scales ranging from 10 2 m to 10 6 m (Dooge, 1988; Gupta & Waymire, 1998a), and provides a natural partitioning of a landscape into hillslopes and links. The mathematical formulation underlying the scaling theory for floods is based on this partitioning. This theory is designed to understand statistical scaling in tenus of Prediction of statistical scaling in peak flows for rainfall-runoff events 101 Fig. 1 A system of hillslopes and links for a drainage basin. physical processes and network geometry across multiple spatial scales of a channel network. Empirical evidence shows that breaks can appear in statistical scaling relationships for annual peak flows in nested basins (Goodrich et al., 1997) as well as in un-nested basins in homogeneous regions (Cathcart, 2001). Although, no breaks in peak flow scaling have been observed in individual rainfall-runoff events (Ogden & Dawdy, 2003), the theory provides a physical framework to understand scaling breaks in floods (Gupta, 2004). Following Gupta & Waymire (1998a), assume that stream flow in a network, T, is observed in multiples of a characteristic time scale At at t = nAt, n = 0, 1,2... One can think of At as the time it takes for water to flow through a link with some mean velocity, v, i.e. At = //v. For simplicity, Gupta & Waymire (1998a) assumed that At was the same for all the links. Many results were derived using this assumption for idealized channel networks (Gupta et al., 1996; Menabde et al., 2001; Menabde & Sivapalan, 2001; Troutman & Over, 2001), but it is too restrictive for natural channel networks in which both link lengths and link velocities vary amongst links. Therefore, following Mantilla et al. (2004), we will take time to be continuous rather than discrete by letting At >0 in the continuity equation given in Gupta & Waymire (1998a). Let q(e,t), a s t , t > 0 be a space-time field representing river discharge, or the volumetric flow per unit time, at time t at the outlet of a link e. Each link, e represents the outlet of a sub network draining into it. Let R(e,t)c(e) denote the volume of runoff into a link e from adjacent hills at time t, where c(e) denotes the area of hills draining into the link e, and R(e,t) is the runoff intensity measured in the units of length per unit time. Finally, let 102 Vijay K. Gupta S(e,t) denote the total volume of runoff stored in a link e at time t, and let dS/dt denote the rate of change of total volume at time t. The equation of continuity for a link representing a control volume for the hillslope-link system can be written as: The sum on the right hand side of Equation (1) consists of discharges from all those links/that join link e at its top. The formulation of the continuity equation does not assume the channel network is binary. However, it assumes that no loops are present in the network. The topology and geometry of a network enters explicitly into the solution of Equation (1) through the summation term as illustrated below through an example. A momentum balance equation is required for expressing the link storage, S(e,t), as a function of discharge, q(e,t); see Reggiani et al. (2001) for a formal derivation of these equations at the hillslope-link scale. The runoff intensity R{e,t) in Equation (1) can be determined from a water balance equation for a hillslope draining into the link e. It includes overland storage, saturated and unsaturated subsurface storage, precipitation, vapotranspiration (ET) and runoff (Gupta & Waymire, 1998a). Spatial hydraulic- geometric relationships are required for completing the mathematical formulation (Menabde & Sivapalan, 2001; Mantilla et al, 2004). The above formulation assumes that the integral dynamical equations governing runoff generation from a hillslope and water transport through a link are available. Both these scales are much larger than the laboratory scales that are typically of the order of 0.1-1.0 m. Individual hydrological processes at the laboratory scale, e.g. flow through saturated and unsaturated columns of a porous medium, are fairly well under- stood. By contrast, the spatial scale for a single hillslope is of the order of 1-10 2 m at which spatial variability becomes important, and new physical elements, e.g. hillslope topography and vegetation, influence the process of runoff generation. The key scale problem here is to spatially integrate laboratory-scale equations over a hillslope, and use these integrated equations for specifying hillslope runoff, R(e,t) in Equation (1). Some basic work has been published on deriving an aggregate behaviour of infiltration on a spatially variable hillslope (Chen et al, 1994). Duffy (1996) has investigated an integral representation of runoff-generation from a hillslope as a dynamic system. However, much more basic theoretical and experimental work remains to be done on this important problem. Momentum equations of fluid mechanics for open-channel flows at the hydro- dynamic scale are well established. They can be used as the starting point for obtaining aggregate flow dynamics for a link. Reggiani et al. (2001) have developed a new framework to write a link-based momentum balance equation, but specific model assumptions are needed to complete the mathematical specification. For example, one problem is to model spatially variable channel boundary roughness within a channel link. Kean (2003) has investigated this problem for modelling boundary sheer stress and flow field near the banks of streams. This type of a theoretical development is required for deriving an average momentum balance equation for a link. In addition, hydraulic-geometric relationships are needed to extend the momentum balance equation to all the links in a network, because velocity, width, depth, slope, and dS(e,t) dt q(e, t) + q (f, t) + R(e, t)c(e), es %,t > 0 (1) Prediction of statistical scaling in peak flows for rainfall-runoff events 103 friction vary spatially over a channel network. For example, Ibbit et al. (1998) give a unique set of field measured spatial hydraulic-geometric relationships on a channel network in New Zealand that clearly exhibit power law relationships for the means but stochastic variations around the means. A derivation of hydraulic-geometric relation- ships from first principles involving a combination of deterministic and stochastic elements remains a major open scale problem. STATISTICAL SCALING AND DYNAMIC PARAMETRIC COMPLEXITY DUE TO SPATIAL VARIABILITY Spatial variability in physical parameters describing runoff-generation processes arises due to differences in geometric, hydraulic and biophysical properties among individual channel links and hillslopes in a drainage basin. The number of hillslopes and links increases with the area of a drainage basin representing spatial scale. Therefore, the number of different values that the physical parameters can take also increases with scale. Gupta (2004) calls this the number of degrees of freedom (NDOF) to underscore the idea that hydrological complexity in a river basin can be compared to the complexity of a statistical-mechanical system that has a very large NDOF (Lienhard, 1964; Gupta & Waymire, 1983). To illustrate that the NDOF in our physical system is very large, let us take the example from Gupta (2004). He considered an idealized bucket-type representation of runoff-generation and transport processes for a single hillslope-link system following Menabde & Sivapalan (2001, pp. 1003-1004). Four physical parameters are required to specify runoff generation for each hillslope. These are, the (j) index governing infiltration and Hortonian runoff generation, a characteristic time scale, T e , governing vapotranspiration, a characteristic time scale, T s , governing subsurface flow, and threshold soil-moisture storage, So. There are three parameters governing link dynamics: the friction coefficient given by Chezy's resistance coefficient, C; the link width, W; and link depth, D. The values of the seven physical parameters vary spatially among hillslopes and links throughout a basin. For illustration assume that a typical hillslope has an area of 0.05 km 2 . Therefore, a 1 lan 2 basin can be partitioned into 1/0.05 = 20 hillslopes and 10 links, because each link is drained by two hillslopes. The number of different spatial values to the seven dynamic parameters is: (20 x 4) + (10 x 3) = 110. Therefore, in a basin of 1 km 2 , NDOF =110. This simple calculation generalizes to an arbitrary size basin of drainage area, A, and leads to the formula: NDOF=UQA (2) It shows that the NDOF increases linearly with the spatial scale of a basin, A. For example, in a small-size basin of 100 km 2 , NDOF = 11 000, and in a medium-size basin of 5000 km 2 , NDOF is about half a million. In order to solve Equation (1) either analytically or numerically, statistical and other simplifying assumptions are necessary to reduce the NDOF. Unfortunately, assumptions regarding spatial variability of these seven parameters cannot be tested directly against empirical observations because these physical parameters cannot be measured across hundreds and thousands of hillslopes. Remote sensing can provide 104 Vijay K. Gupta some of the spatially variable information, especially as it pertains to channel network geometry, precipitation and ET. Still, most of the parametric information regarding the processes of runoff generation and transport for spatially variable hillslopes and links is not available. The simple calculation given in Equation (2) illustrates the funda- mental difficulty that underlies distributed rainfall-runoff modelling due to dynamic parametric complexity. In the absence of a theoretical framework, hydrologists and engineers have routinely and widely used calibration of dynamic parameters to reduce large NDOF. But calibration is an ad hoc approach and is unsuitable for developing a foundational understanding of space-time rainfall-runoff relationships (Klemes, 1997; Woolhiser, 1996). Statistical scaling provides a new theoretical framework to reduce the large NDOF of dynamic parameters as explained below. In contrast to the space-time complexity that physical processes exhibit at the hillslope-link scale, empirical observations show that the aggregated behaviour of peak flow statistics in rainfall-runoff events exhibits statistical scale invariance or scaling at successively larger spatial scales of a drainage network (Ogden & Dawdy, 2003). Scaling is an asymptotic property, which holds in the limit of large area. This result was obtained analytically for the idealized Peano network model by Gupta et al. (1996). Troutman & Over (2001) generalized it to the class of mean self-similar networks. Menabde et al. (2001) considered attenuation in flow by taking dS/dt > 0 in Equation (1) on a Peano network, and found using numerical calculations that scaling holds asymptotically for large areas. Because scaling is an asymptotic property, the scaling parameters must remain insensitive to many physical details of the system at the hillslope-link scale. This is the key idea that allows us to reduce the large NDOF of dynamic parameters in the present context. This situation is somewhat similar to the Central Limit Theorem (CLT) of probability theory, in which the convergence of a random sum to a normal distribution does not depend on all the details of the individual probability distributions of the summands, but on only the common mean and variance of individual summands. However, demonstration of asymptotic scaling in peak flows on real channel networks under both idealized and realistic flow conditions remains a very important open theoretical problem. Spatial statistical scaling in peak flows refers to similitude or similarity across multiple scales and is represented by log-log linear equations between the mean, or higher order finite statistical moments, and the drainage area, which serves as the scale parameter. The slopes and the intercepts of log-log linear relationships are called scaling parameters. They represent integrated basin-scale behaviour. Scaling can be defined in terms of quantiles instead of moments. This is particularly important for extreme events whose probability distributions often exhibit "fat" tails, and a priori one does not know the order above which the statistical moments are infinite. The scaling of quantiles can be used to empirically estimate an upper bound for the order of finite moments (Pavlopoulos & Gupta, 2003). Scaling parameters are analogous to macroscopic laboratory-scale parameters, e.g. hydraulic conductivity, fluid viscosity and density, in so far as they arise from microscopic statistical-dynamical equations governing molecular motions with very large NDOF. The goal of the scaling theory of floods on river networks is to derive scaling parameters from physical processes on real channel networks, which is similar to the goal of deriving macroscopic parameters from microscopic statistical physics. Prediction of statistical scaling in peak flows for rainfall-runoff events 105 TESTS OF PHYSICAL ASSUMPTIONS WITHIN A STATISTICAL SCALING FRAMEWORK Scaling theory predicts scaling parameters either numerically or analytically through solutions of Equation (1). These predictions can be compared with empirical estimates of the parameters for testing different physical hypotheses. Testability gives a new framework for understanding the spatially variable physics of rainfall-runoff processes as illustrated below through a concrete but idealized real world example. My main objective is to demonstrate within the context of the scaling framework how predict- ions based on a set of idealized assumptions can be tested against data. Falsification of model predictions paves a way to formulating and testing new physical assumptions that can capture more of the essential characteristics of real systems. Tests of model predictions also provide a basis for conducting new experiments and observations. In due course, a continual interaction among theory, data and modelling leads to new scientific understanding, as illustrated by Klemes (1997, Fig. 1(a)). The reader should contrast this iterative logic that is designed to provide a greater understanding of the physical system, with the logic of improving the fit of a given model to data using calibration of model parameters, which does not lead to an improved understanding of the system (Klemes (1997, Fig. 1(b)). Consider the solution of Equation (1) under three assumptions: (a) change in storage in each link, dS/dt = 0; (b) runoff intensity per unit area into each link is given by: R{e,t) = 1, t = 0 and R(e,t) = 0, t > 1 ; (c) hillslope areas draining into each link e are the same, c(e) = 1. It should be noted that the second assumption is the same as one of the key assumptions underlying the theory of IUH. Substituting these three assumptions into Equation (1) gives: f The solution of Equation (3) under a further assumption that the velocity of flow on the channel network, v, is constant in time and space is given by: where W e {x), x > 0, is the local geomorphic width function of a network. It is defined as the number of links at a flow distance x from the bottom of a link e representing the outlet of a sub-basin draining into it (Gupta & Waymire, 1998a). It is a spatial general- ization of the width function introduced by Kirkby (1984) and Lee & Delleur (1976), which served as the basis for developing Gomorphologie IUH (GIUH) theory based in the width function (Gupta & Waymire, 1983; Troutman & Karlinger, 1984; Gupta & Mesa, 1988). The spatial field of floods over an arbitrary time interval at the bottom of a link, e, is given by taking the maximum on both sides of Equation (4) over that time interval: Q{e) = maxq(e,t) = maxW e (vt) = Q(e), eex (5) We will assume in the following example that the time interval in Equation (5) corresponds to that of an individual rainfall-runoff event. The local geomorphic width function of a natural channel network varies stochast- ically from one link to the next due to spatial variability in the branching and the q(e,t) = Y i q(f,t); t>0,e,fez (3) q(e,t) = W e (vt),t>0,e,f e % (4) 106 Vijay K. Gupta geometric structures of a network. Therefore, the maxima of the local width functions and the peak flows in (5) represent spatial random fields. Troutman & Karlinger (1984) were the first to consider the spatial dependence of width function for the random model of channel networks. Spatial dependence of floods and width functions is a fundamental aspect of the scaling theory of floods. It is quite different from the GIUH theories and rainfall-runoff models in which space is fixed at a basin outlet (Rodriguez-Iturbe & Rinaldo,1997). I will illustrate this key idea in the example given below. The dependence of runoff on spatial scales of a channel network arises through the indexing parameter, e, denoting a link. However, spatial scales for a channel network are not unique and any one of several choices is possible. As a simple example, consider the spatial scale given by the drainage area, Aie), draining into the link e. Since each sub-basin draining into a link has a different area, it follows that the field of peak flows, Q(e), and the runoff field, q(e,t), depends on the multiple spatial scales of sub-basins within a basin. From here on we will denote the dependence of the peak flow and maximum of the width function on drainage area and write, Q(e) = Q(A(e)) = Q(A), and (e) = (A(e)) = (A). Therefore, Equation (5) can be rewritten as: Q(A) = Q(A),VA (6) We computed width function maxima for the Goodwin Creek basin using a GIS software called hidrosig. It has been developed by my research group to conduct state of the art hydrological analyses and numerical simulations on a river basin after partitioning it into hillslopes and links as shown in Fig. 1. The empirical estimation of width function maxima needs an explanation of some key issues that arise in this context. First, note that total drainage area, A(e), is the sum of all the hillslope areas draining into a link e. Summing over all the hillslope areas, it follows that, A(e) - c(2m{e) -1), where m(e) is the link magnitude representing the total number of source links draining into the link e, (2m(e) -1) is total number of links and c is the mean hillslope area (Shreve, 1967). From our previous assumption that hillslope areas draining into each link, c(e) = 1, it follows that c = 1. Magnitude and drainage area vary stochastically from one link to the other throughout a network. Therefore, they are not always the most convenient scale parameters. As an alternative, I consider another well-known scale parameter, known as the Horton order, denoted by ((e), or simply co. Horton order varies from 1 to 4 on the Goodwin Creek basin. We take all links of the same Horton order and compute the mean magnitude, m , and the mean of the width function maxima, 0 (m), for these links. Mean magnitude for a fixed order serves as the scale parameter in this computation. Figure 2 shows the relationship between the empirical, 0 (m), and the mean magnitude m. Each point represents a given Horton order, and the order increases from 1 to 4 going from left to right. The log-log linearity holds well and a regression equation fit to it can be written as: log E[Q 0 (m)] = y log m + b (7) The Goodwin Creek data gives an estimate of the slope, y = 0.46 as shown in Fig. 2. Technical issues arise in estimating the scaling parameter, y, because regression is not Prediction of statistical scaling in peak flows for rainfall-runoff events 107 1001 10 100 1000 Mean Magnitude (m) Fig. 2 Scaling of mean width function maxima with respect to mean magnitude for Goodwin Creek basin, Mississippi. quite suitable for this purpose. Mantilla et al. (2004) have explained some of these estimation issues. I use regression only for illustration. Equations (6) and (7) lead to two theoretical predictions: (a) that peak flows exhibit statistical scaling, and (b) the flood scaling exponent 8 is given by: DATA NEEDS TO TEST DISTRIBUTED PHYSICAL ASSUMPTIONS USING SCALING PREDICTIONS Empirical tests of model predictions described above require that observed streamflow hydrographs are available for individual rainfall-runoff events at many spatial locations on a channel network. This type of streamflow information at several locations is rarely available to test predictions of scaling theory. Goodwin Creek basin is an exception because it contains 14 streamflow gauges in a 21 lan 2 area. This allowed us to test the two scaling predictions given above. Empirical observations of peak flows for individual rainfall-runoff events on the Goodwin Creek basin have shown that they exhibit spatial statistical scaling (Ogden & Dawdy, 2003). For example, Fig. 3 shows a plot of observed peak flows against drainage areas at 14 stteam- flow gauges on the Goodwin Creek basin for the rainfall-runoff event of 12 March 1986. A fitted regression equation can be interpreted as scaling in mean peak flows conditioned on area and can be written as: Here, the mean peak flow at some reference drainage area is denoted by a. Equation (9) shows that log-log linearity, or a power law, holds between mean peak flows and the drainage areas predicted by Equations (6) and (7). Figure 3 gives an empirical value of the flood scaling exponent, 9 = 0.82. The standard regression analysis is used here mainly for illustration. Interesting technical issues arise in estimating scaling exponent without using regression that have been investigated elsewhere (Mantilla et ai, 2004). G = y = 0.46 (8) io g [e 0 (^)] = e i o g ^ + f (9) 108 Vijay K. Gupta 100. 000 So 10. 000 1 1.000 2 0. 100 s ^ 0. 010 0. 001 0.1 1.0 10.0 Area Fig. 3 Scaling of observed peak flows for the event of 12 March 1986, Goodwin Creek. The observed value of the flood-scaling exponent does not agree with the predicted value of 0. 46 given by Equation (8). This failure should not come as a big surprise to the reader because the set of three physical assumptions used here for making this prediction are highly idealized. Nonetheless, this example demonstrates how the physical assumptions can be tested within the scaling framework. I wish to make a major point regarding streamgauging with this example. Tests of the scaling theory of floods in other basins in the USA and other countries will need streamflow data at several locations on nested basins for individual rainfall-runoff events. Such data sets are rarely available either in the USA or other countries because streamflow measurements requiring the establishment of standard USGS gauging stations are too expensive or impractical. Under these circumstances, emplacement of pressure gauges at judiciously chosen field sites and conversion of the stage measure- ments to discharge with a rating curve generated from an appropriately designed flow model and ancillary models, is a viable and reasonably accurate choice. In order to be useful for generating discharge-rating curves, a flow model must be able to treat curved channels with irregular boundaries without employing any empirically adjusted parameters. Consequently, the model must explicitly calculate the effective flow resistance as a function of stage from measurements of: (a) the physical roughness of the river bed, (b) the physical roughness of the channel banks, (c) the physical characteristics of the bank vegetation, (d) the physical roughness of the flood plain, and (e) the physical characteristics of the flood plain vegetation. We are currently developing and verifying flow, sediment transport, and geomorphic adjustment models for this purpose using the method of Kean ( 2 003 ) . In the long run, simultaneous observations of space-time rainfall, streamflows, and vapotranspiration will be needed within individual mesoscale river basins for conducting scaling analyses of floods and other hydrological variables. Such data sets are typically unavailable in the USA or the rest of the world. Substantial progress on these scaling and other problems in mesoscale basins requires that new comprehensive field programmes be undertaken in the next decade. A recent report to the National Science Foundation, USA (http://cires.colorado.edu/hydrology) articulated this key hydrological issue. This issue is also fundamental for developing and testing new theories and models for PUB (Sivapalan et al, 2 003 ) . Prediction of statistical scaling in peak flows for rainfall-runoff events 109 FUTURE RESEARCH How do we relax some of the physical idealizations that led to predicting the flood scaling exponent in Equation (8)? This is a big question that I will touch upon very briefly. The empirical results given by Ogden & Dawdy (2003) show that the mean flood scaling exponents for over 200 individual rainfall-runoff events vary between 0.9 and 0.6 with a mean of about 0.8. Therefore, the flood scaling exponents are significantly larger than 0.46 predicted by the width function maxima. By contrast, the simulation results given by Menabde & Sivapalan (2001) and Menabde et al. (2001) for idealized mean self-similar channel networks showed that incorporation of more realistic link dynamics than contained in our three assumptions here, produced a flood scaling exponent that is below the value predicted by the scaling of the width function maxima. What are the physical reasons that cause the empirical flood scaling exponent to be larger than the width function scaling exponent on the Goodwin Creek and the Walnut Gulch basins? How can we understand the physical basis of variability in the flood scaling exponents in individual rainfall-runoff events? My current work at the University of Colorado is investigating this issue by incorporating space-time variable rainfall into this mathematical framework for testing flood scaling predictions for the Goodwin Creek basin (Furey & Gupta, 2004). Likewise, event-based flood scaling analyses on the Walnut Gulch basin, Arizona, is giving new physical insights into how the spatially variable friction as a hydraulic-geometric variable affects the flood scaling exponent (Mantilla et al, 2004). Acknowledgements Some of the ideas in this paper were presented at a conference on Water and Environment (Bhopal, India, 2003), and published in the conference proc- eedings without peer reviews. V. P. Singh, the conference convener, kindly provided comments on an earlier draft of this paper. Y. Alila, University of British Colombia, Canada, sent me several key references, which were extremely useful, reviewed an earlier draft of this paper and provided critical comments on it. Brent Troutman, USGS, Lakewood, Colorado, provided comments on an earlier draft. S. Veitzer, P. Furey, and Ricardo Mantilla, members of my research group gave many insightful comments. Discussions with Jim Smith and Jason Kean, USGS, Boulder, Colorado, on this research have been very fruitful. These comments helped me to clarify important concepts and they resulted in a substantial improvement of this paper. I am grateful for all their input. This research was supported by grants from NSF and NASA. REFERENCES Barenbl at t , G. I. ( 1996) Scaling, Self-Similarity and Intermediate Asymptotics. Cambr i dge Text s in Appl i ed Mat hemat i cs 14. Cambr i dge Uni versi t y Pr ess, Cambr i dge, UK. Bescht a, R. L., Pyl es, M. R., Skaugset , A. E. & Surfleel, C. G. ( 2000) Peakfl ow r esponse to forest pract i ces in the west ern Cascades of Or egon, USA. J. Hydrol. 233, 102- 120. Cat hcart , .1. ( 2001) The effects of scal e and st orm severi t y on the linearity of wat er shed r esponse reveal ed t hrough t he regi onal L- moment anal ysi s of annual peak flows. PhD Di ssert at i on, Uni versi t y of Bri t i sh Col ombi a, Vancouver , Canada. Chen, Z. Q. , Govi ndaraj u, R. S. & Kavvas, M. L. ( 1994) Spatial aver agi ng of unsat urat ed flow equat i on under infiltration condi t i ons over areal l y het er ogeneous condi t i ons, 1 and 2. Water Resour. Res. 30( 2) , 523- 535. 110 Vijay K. Gupta Dooge, J. ( 1988) Hydr ol ogy in perspect i ve. Hydrologie Sci. J. 33, 61- 85. Duffy, C. A. ( 1996) Two state i nt egral -bal ance model for soil moi st ur e and gr oundwat er dynami cs in compl ex terrain. Water Rsout: Res: 32 ( 8) , 2 42 1- 2 43 4. Furey, P. & Gupt a, V. K. ( 2004) Test s of physi cal hypot heses regardi ng rainfall and local wi dt h function for predi ct i ng flood-scaling exponent on Goodwi n Cr eek basi n, Mi ssi ssi ppi , USA (prepri nt ). Goodr i ch, D. , Lane, L. J. , Shi l l sl opei t o, R. M. , Miler, S. N. , Syed, K. H. & Wool hi ser, D. A. ( 1997) Li neari t y of basi n r esponse as a function of scal e in a semi ari d wat er shed. Water Rsout: Res. 33( 12 ) , 2 951- 2 965. Gupt a, V. K. ( 2004) Emer gence of statistical scal i ng in floods on channel net wor ks from compl ex runoff dynami cs. Chaos, Solitom and Fractals 19, 3 57 - 3 65. Gupt a, V. K. & Dawdy, D. R. ( 1995) Physical i nt erpret at i ons of regional vari at i ons in t he scal i ng exponent s of flood quant i l es. Hydrol. Processes 9( 3/4) , 3 4 7 - 3 6 1 . Gupt a, V. K. & Mesa, O. ( 1988) Runoff generat i on and hydrol ogi e response via channel net wor k geomor phol ogy: recent progress and open pr obl ems. J. Hydrol. 10 2 , 3 - 2 8. Gupt a, V. K. & Waymi r e, E. C. ( 1983) On t he formul at i on of an anal yt i cal appr oach to hydrol ogi e response and si mi l ari t y at t he basi n scal e. J. Hydro/. 65, 95- 12 3 . Gupt a, V. K. & Waymi r e, E. C. (1998a) Spatial vari abi l i t y and scal e i nvari ance in hydrol ogi e rgi onal i sat i on. In: Scale Dependence and Scale Invariance in Hydrology (ed. by G. Sposi t o), 88- 13 5. Cambr i dge Uni versi t y Press, Cambr i dge, UK. Gupt a, V. K. & Waymi r e, E. C. ( 1998b) Some mat hemat i cal aspect s of rainfall, l andl br ms and floods. Stochastic Methods in Hydrology: Rainfall, Landforms und Floods (ed. by O. E. Barndorff-Ni el sen, V. K. Gupt a, V. Per ez- Abr eu & E. C. Waymi r e) , 12 9- 17 1. Advanced Seri es on Statistical Sci ences and Appl i ed Probabi l i t y, V7. Worl d Scientific, Si ngapore, pp. 207. Gupt a, V. K., Mesa, O. & Dawdy, D. R. ( 1994) Mul t i scal i ng t heory of flood peaks: regi onal quant i l e anal ysi s. Water Rsout: Res. 30 ( 12), 3 405- 3 42 1. Gupt a, V. K., Cast r o, S. & Over, T. M. ( 1996) On scalinsx exponent s of spatial peak flows from rainfall and river net work geomet r y. J. Hydrol. 18 7 , 81- 104. Ibbilt, R. P., McKer char , A. 1. & Duncan, M. .1. ( 1998) Taieri Ri ver data t o test channel net wor k and river basin het erogenei t y concept s. Water Rsout: Res. 34 ( 8) , 2 085- 2 088. Kean, J. ( 2003) Comput at i on of flow and boundar y sheer st ress near t he banks of st reams and ri vers. PhD Thesi s, Uni versi t y of Col or ado, Boul der, Col or ado, USA. Ki rkby, M. ( 1984) Test s of r andom net wor k model and its appl i cat i on to basin hydr ol ogy. In: River Networks (ed. by R. S. Jarvi s & M. J. Wol denber g) , 3 55- 3 68. Benchmar k Papers in Geol ogy 80. Hut chi nson Ross, St roudsburg, PA (Ori gi nal l y publ i shed in 1976). Kl emes, V. ( 1997) Of carts and horses in hydrol ogi e model i ng. J. Hydrol. Engng 2,43-49. Lee, M. T. & Delleur, J. W. ( 1976) A vari abl e source area model of t he rai nfal l -runoff pr ocess based on t he wat er shed st ream net wor k. Water Rsout: Res. 12, 1029- 1035. Li enhard, .1. II. ( 1964) A st at i st i cal -mechani cal predi ct i on of t he di mensi onl ess unit hydr ogr aph. J. Geophys. Res. 69( 24) , 52 3 1- 52 3 8. Mandel brot , B. (1982) The Fractal Geometry of Nature. Freeman, San Franci sco, USA. Mant i l l a, R., Gupt a, V. K. & Mesa, O. J. (2004) Test of physi cal hypot heses r egar di ng channel friction for predi ct i ng fl ood-scal i ng exponent on Wal nut Gul ch Basi n, Ar i zona, USA. J. Hydrol. (under revi ew). Menabde, M. , Vei t zer, S., Gupt a, V. K. & Si vapal an, M. ( 2001) Test s of peak flow scal i ng in si mul at ed sel f-si mi l ar river net wor ks. Adv. Water Rsout: 2 4 , 991- 999. Menabde, M. & Si vapal an, M. ( 2001) Li nki ng s pace- t i me variability of rainfall and runoff fields: a dynami c appr oach. Adv. Water Rsout: 2 4 , 1001- 1014. Ogden, F. & Dawdy, D. R. ( 2003) Peak di schar ge seal i ng in a smal l Hort oni an wat ershed. J. Hydrologie Engng 8 , 6 4 - 7 3 . Pavl opoul os, II. & Gupt a, V. K. ( 2003) Scal e i nvari ance of regi onal wet and dry durat i ons of r ai nf i el ds: a di agnost i c st udy. J. Geophys. Res. 10 8 ( D8) , 83 - 87 . Pol l , N. L., Al l an, J. D. , Bai n. M. B. , Karr, J. R., Prest ergaard, L. , Ri cht er, B. D. , Spar ks, R. E. & St r ombcr g, J. C. (1997) The nat ural flow regi me. Bioscience 4 7 , 769- 784. Reggi ani , P., Si vapal an, M. , 1 l assani zadeh, S. M. & Gray, W. G. ( 2001) Coupl ed equat i ons for mass and moment um bal ance in a st ream net wor k: t heoret i cal deri vat i on and comput at i onal exper i ment s. Proc Rov. Soc London A 4 57 , 157- 189. Rodr i guez- I t ur be, 1. & Ri nal do, A. ( 1997) Fractal River Basins. Cambr i dge Uni versi t y Press, Cambr i dge, UK. Shreve, R. L. ( 1966) Statistical l aw of st ream number s. . / . Geol. 7 4 , 17- 37. Shr eve, R. L. ( 1967) Infinite t opol ogi cal l y r andom channel net wor ks. J. Geol. 7 5, 178- 186. Si ngh, V. P. ( 1995) Computer Models of Watershed Hydrology. Wat er Resour ces Publ i cat i ons, Hi ghl ands Ranch, Col or ado, USA. Si vapal an, M. , Takeuchi , K., Franks, S., Gupt a, V., Kar ambi r i , IL, Lakshmi , V. , Li ane, X., McDonnel l , .1., Mendi ondo, E., O' Connel l , P. , Oki , T., Pomer oy. .1., Schert zer. D. , Uhl enbr ook, S. & Zehe, E. ( 2003) I AHS decade on predi ct i ons in ungauged basi ns ( PUB) , 2 003 - 2 012 : Shapi ng an exci t i ng future for the hydrol ogi e sci ences. Hydrol. Sci. J. 48(6), 857- 880. Smi t h, ,1. ( 1992) Represent at i on of basi n scal e in flood peak di st ri but i ons. Water Rsout: Res. 28 ( 11 ), 2 993 - 2 999. Prediction of statistical scaling in peak flows for rainfall-runoff events 111 Sposi t o, G. ( 1998) Scale Dependence and Scale Invariance In Hydrology. Cambr i dge Uni versi t y Press, Cambr i dge, UK. Tr out man, B. & Karl i nger, M. ( 1984) On t he expect ed wi dt h function of t opol ogi cal l y r andom channel net wor ks. J. Appl. Prob. 22, 836- 849. Tr out man, B. & Over, T. ( 2001) Ri ver flow mass exponent s with fractal channel net wor ks and rainfall. Adv. Water Resour. 2 4 , 967- 989. Turcot t e, D. ( 1997) Fractals and Chaos in Geology and Geophysics (2nd edit). Cambr i dge Uni versi t y Press, Cambr i dge, UK. Vei t zer, S. & Gupt a, V. K. ( 2001) Statistical self-similarity of wi dt h-funct i on maxi ma with i mpl i cat i ons for floods. Adv. Water Resour. 2 4 , 955- 965. Vei t zer, S. & Gupt a, V. ( 2000) Random sel f-si mi l ar river net wor ks and deri vat i ons of general i zed Hort on l aws in t erms of statistical si mpl e scal i ng. Water Resour. Res. 36( 4) , 1033- 1048. Wol man M. G. & Mi l l er, G. P. ( 1960) Magni t ude and frequency of forces in geomor phi c processes. J. Geol. 68, 54- 74. Wool hi ser, D. A. ( 1996) Search for physi cal l y based runoff model a hvdrol ogi c El Dor ado? J. Hydrant Fngng ASCE 1 2 2 ( 9) , 122- 129.