Вы находитесь на странице: 1из 3

FORUM

Prediction versus Understanding


„The 2006 Ven Te Chow Lecture…
David R. Dawdy flow hydrographs and peak flows on the time scales of individual
3055 23rd Ave., San Francisco, CA 94132. rainfall-runoff events. This in contrast to the statistical approach,
Downloaded from ascelibrary.org by UNIVERSITY OF VIRGINIA on 11/03/13. Copyright ASCE. For personal use only; all rights reserved.

which was spatial. The two approaches were developed to solve


The times they are a-changin’. Global warming is upon us. Engi- practical engineering problems. The regional statistical analysis of
neering is about prediction. However, in water resources engi- annual flood frequencies was designed to predict floods of a given
neering, many of the prediction tools soon may not be valid. They return period for highway, bridge, and spillway design at un-
may not predict correctly under the new conditions. It may not gauged locations, and rainfall-runoff models were developed for
happen in my lifetime but it certainly will in many of yours. So planning and project design. The dichotomy in the two ap-
what can we do about it? Let me give a little of my personal proaches to hydrologic analyses continues until today in hydro-
history to try to help answer the question I just raised. logic research, education, and practice.
From 1956 to 1958 I worked with Manuel Benson in develop- It was also recognized in the early 1960s that the two ap-
ing the regression method for predicting flood frequencies. The proaches must be closely connected because physical rainfall-
method uses historical streamflow records to predict flood fre- runoff mechanisms generating floods were the basis for the flood
quencies at ungauged basins in space. That is the method used by values used in the statistical analysis. For example, federal agen-
the USGS and elsewhere today. In 1958, I worked for the summer cies other than the USGS have often used deterministic rainfall-
with Walter Langbein. Walter told me that the Ground Water staff runoff analyses to determine flood frequencies. An intercompari-
wanted to do the prediction of the low flows for the Potomac son of use of the various models for the prediction of flood
River for use in water supply planning by the District of Colum- frequencies was undertaken, and it was shown that the USGS
bia. The Surface Water staff predicted the flows by use of a re- model predicted with less bias and less variance than any of the
gression model. Their method was black box. The Ground Water other models. Thus, once again, better understanding did not lead
group wanted to add understanding to the prediction to make it to better prediction. However, a general theoretical framework for
better. After all, low flows are made up from water draining from connecting the two, i.e., how the physics of the system could be
aquifers and the equations for those flows were well known. used to predict spatial statistical parameters in quantile analysis
Walter said this happened every 5 years or so, and no one ever was not known, and tools had not been developed to address this
came up with a better predictor than the black box regression fundamental hydrologic question. The parameters used in the re-
model. So, understanding does not necessarily lead to better pre- gression models were surrogates for the physics underpinning the
diction. Prediction is an engineering problem. Understanding, in prediction. However, it was not recognized at the time that the
general, is a scientific problem. problem of connecting the two approaches was a 共nonlinear兲 geo-
Some years back, Peter Eagleson chaired an NRC Committee, physical problem belonging to hydrological sciences rather than a
Opportunities in the Hydrologic Sciences, which produced a pub- practical engineering problem belonging to hydrologic engineer-
lication titled Opportunities in the Hydrologic Sciences 共WSTB ing and that it required the development of new ideas and new
1991兲. That report called for developing hydrology as a geo- tools.
science. It outlined several steps toward achieving that goal. Statistical analysis has not been confined to regression analy-
However, somehow there was no immediate follow through. Hy- sis. The Harvard Group 共Fiering 1964兲 introduced stochastic
drology grew out of engineering and was mainly concerned with simulation, and a plethora of papers followed purporting to dem-
prediction. What that committee recommended was that the hy- onstrate how stochastic simulation of streamflows could aid in
drologic community should begin to stress the science behind project design. Mandlebrot and Wallis 共1968兲 introduced their
hydrology and concentrate more on understanding the water cycle model of the Hurst effect in 1968. This introduced self-similarity
as a whole. Rather than follow the lead of the practical problems 共which produces power laws or scaling兲 and long memory pro-
of the several water agencies, the hydrologic community should cesses into the hydrologic tool chest. Attempts to model the Hurst
define what the present problems were and future problems were effect with different models followed, culminating in the intro-
to be and should cooperate to attack those problems as a scientific duction of the broken-line model 共Rodriguez-Iturbe et al. 1972兲.
challenge. One of the problems they touched on was the effect of All these models were concerned with time series; whereas, ear-
greenhouse gases on global hydrology. lier regression models were concerned with spatial relationships,
There was a large background of history leading up to the i.e., they could be used to predict floods of given frequencies at a
Committee and its conclusions. I will give one example from my point or points in space within a “homogeneous region.” The
direct experience. The USGS developed the statistical approach to meaning of long memory 共a power law correlation function which
the prediction of flood frequencies as represented by the regional has infinite memory in most models of the Hurst effect兲 was never
共spatial兲 quantile analysis of annual flood frequencies mentioned explained, although it was demonstrated that it applied to the
earlier. The Stanford watershed model 共1964兲, and subsequently modeling of streams in a karst area 共Garcia et al. 1972兲.
other rainfall-runoff models, developed the deterministic ap- The study of long memory processes and of self-similar
proach based on semiempirical physical equations. It described processes died down without any physical understanding or ex-

JOURNAL OF HYDROLOGIC ENGINEERING © ASCE / JANUARY/FEBRUARY 2007 / 1

J. Hydrol. Eng. 2007.12:1-3.


planation of the theory behind the processes and their true impli- Walnut Gulch data analysis of Goodrich et al. 共1997兲 exhibits
cation for understanding hydrology. Nevertheless, the fundamen- a break in the flood frequency power law relations called simple
tal question was raised in the open literature 共Klemes 1974兲. In scaling 共Gupta and Waymire 1998兲. The break occurred at an area
the concluding remarks to my Kisiel Lecture at the University of of about a quarter of a square mile. Above the break point, the
Arizona 共Dawdy 1995兲, I stated, “Some one is going to…end up slopes of the flood frequency relations are close to a half. A physi-
with an understanding of the relation of the physical basis for cal basis of this slope value is related to the geometry and dy-
statistical variability in time and space.” I still believe that and namics of flow in a network 共Mantilla et al. 2006兲. Gupta and
will illustrate this key issue below. Waymire 共1998兲 offered a physical explanation of similar values
I mentioned earlier that I was research assistant to Manuel of power law slopes for two regions of New Mexico that were
Benson in the development of the USGS flood frequency meth- observed in the USGS annual flood frequency analysis 共Gupta
ods. The USGS applied those models on a state-by-state basis. I and Dawdy 1995兲. Goodrich et al. 共1997兲 explained their break as
always thought there was more information in those reports than the result of the influence of the size of storm cells in their basin.
Downloaded from ascelibrary.org by UNIVERSITY OF VIRGINIA on 11/03/13. Copyright ASCE. For personal use only; all rights reserved.

was utilized. Each report used data outside its immediate region For basins smaller than the cell size, the flood frequencies in-
to determine their equations, but there was no real attempt to creased almost directly with area. Ogden and Dawdy showed that
coordinate the results. Could there be a universal equation cover- the difference from a direct relation was related to the actual path
ing the United States? How could the information in those reports of the storm and whether, in general, the storm hit directly over
be used other than to develop prediction equations region by re- the basin or the storm center was outside the basin. If every storm
gion? I collected all the reports as they came out, but I never were centered on a basin, seemingly, the flood frequencies should
managed to solve my self-defined puzzle. vary directly with area. Furey and Gupta 共2005兲 showed that tem-
Vijay Gupta was a student in some of my classes at Colorado poral varibility was the cause, which is similar to Ogden and
State University, and I influenced him to go to the University of Dawdy’s 共2003兲 empirical conclusion. Incidentally, the Walnut
Arizona to study under Chester Kisiel. I served on his M.S. thesis Gulch data exhibit simple scaling both before and after the break
and Ph.D. dissertation committees and have followed his career in the flood frequency relations.
through the years. Whenever I visited Vijay at Arizona or, later, at An important conclusion concerning understanding aiding in
the University of Mississippi and the University of Colorado, he prediction has to do with climate change as a result of global
filled me in on his work, which turned more to the understanding warming. By understanding how basins operate physically and
of the influence of basin morphology on the hydrology of the how the physics is connected to the spatial statistical variability in
basin. Finally one day when Vijay was explaining his thinking on floods, we may have a leg up on the influences of climate change
self-similarity of floods 共based in power laws or scaling兲 in joint and other impacts due to natural and human influences. Will glo-
collaboration with a mathematician, Ed Waymire 共Gupta and bal warming change the structure of rainfall intensity and dura-
Waymire 1990兲, I realized that this may be the solution to my tion? What will that change be? If not, there will continue to be a
puzzle. Perhaps the information latent in the USGS flood fre- break at about a quarter of a square mile in the Walnut Gulch
quency relations could be revealed through developing a scaling basin. As global warming changes the frequency of thunderstorms
theory. Those data could be used to test some of the scaling hy- or total precipitation, the return period of the floods at the break
potheses. The result was a series of papers showing a signature of point may change, but the basic structure of the flood frequency
power laws in the USGS flood frequency analyses and an attempt power law relations should remain intact. Similar conclusions
to give a crude physical interpretation to the scaling results may be reached elsewhere. Scaling properties are related to the
共Dawdy and Gupta 1992; Gupta et al. 1994; Gupta and Dawdy physics of the basin and should be invariant.
1995兲. However, it was soon realized that a physical understand- In conclusion, scaling theory at this time is leading to under-
ing of statistical scaling of floods, strictly applied, was valid only standing, a long-standing need in hydrology. It is contributing to
for nested basins, that is, basins with subbasins nested within, and the scientific understanding of hydrology. However, with global
time scales of individual rainfall-runoff events 共Gupta et al. warming introducing nonstationarity into hydrology, the tradi-
1996兲. Although the USGS regions contained similar basins, they tional tools of prediction will no longer be valid. They depend on
were not nested basins, and the time scale was annual not event- the future repeating the past. The future will not. Perhaps scaling
based. The observational evidence for power laws in quantile re- theory or other scientific explanations of hydrology can aid in
gression analysis of annual floods in nested basins for regional or hydrologic prediction in our present hydrologic universe, which is
spatial analysis of annual flood frequencies is a recent develop- becoming nonstationary. This would be a case where understand-
ment, because only two papers have been published on it: one for ing may, in fact, aid in prediction. Such theories may be necessary
the Walnut Gulch basin, Ariz. 共Goodrich et al. 1997兲 and another in our future hydrology.
for the Goodwin Creek basin, Miss. 共Ogden and Dawdy 2003兲.
The USDA has operated these two basins as experimental facili-
ties for many years, and they are quite unique as natural labora- References
tories in terms of high density of rainfall and streamflow data. The
Walnut Gulch basin is in the semiarid southwest United States, Dawdy, D. R. 共1995兲. “Hurst, scaling, and the meaning of hydrology.”
and the Goodwin Creek is in the humid southeast United States. Fourteenth Chester C. Kisiel Memorial Lecture, Univ. of Arizona,
Ogden and Dawdy 共2003兲 also conducted the first event-based Tucson, Ariz.
analysis of spatial flood statistics for the Goodwin Creek basin Dawdy, D. R., and Gupta, V. K. 共1992兲. “Regional flood frequency rela-
tions as evidence of scaling invariance.” Paper presented at Fall Con-
and observed a presence of power laws. Surprisingly, they found
ference of Amer. Geophys. Union, AGU, Washington, D.C.
that the power law parameters vary from one event to the next. Fiering, M. B. 共1964兲. “Multivariate techniques for synthetic hydrology.”
Following on this work, Furey and Gupta 共2005兲 have tested the J. Hydr. Div., 90共5兲, 43–60.
hypothesis that the observed variability in the power law param- Furey, P., and Gupta, V. 共2005兲. “Effects of excess rainfall on the tempo-
eters can be attributed to temporal variability in effective rainfall ral variability of peak-discharge power laws.” Adv. Water Resour., 28,
intensity and duration among events 1240–1253.

2 / JOURNAL OF HYDROLOGIC ENGINEERING © ASCE / JANUARY/FEBRUARY 2007

J. Hydrol. Eng. 2007.12:1-3.


Garcia, L. E., Dawdy, D. R., and Mejia, J. M. 共1972兲. “Long memory Gupta, V. K., and Waymire, E. 共1998兲. “Spatial variability and scale in-
monthly streamflow simulation by a broken-line model.” Water Re- variance in hydrologic regionalization.” Scale dependence and scale
sour. Res., 8共4兲, 1100–1105. invariance in hydrology, G. Sposito, ed., Cambridge University Press,
Goodrich, D. C., Lane, L. J., Shillito, R. M., Miller, S. N., Syed, K. H., New York, 88–135.
and Woolhiser, D. A. 共1997兲. “Linearity of basin response as a func- Klemes, V. 共1974兲. “The Hurst phenomenon: A puzzle?” Water Resour.
tion of scale in a semiarid watershed.” Water Resour. Res., 33共12兲, Res., 10共1兲, 675–688.
2951–2965. Mandelbrot, B. B., and Wallis, J. R. 共1968兲. “Noah Joseph, and Opera-
Gupta, V. K., Castro, S., and Over, T. M. 共1996兲. “On scaling exponents tional Hydrology.” Water Resour. Res., 4共5兲, 909–918.
of spatial peak flows from rainfall and river network geometry.” J. Mantilla, R., Gupta, V., and Mesa, O. 共2006兲. “Role of coupled flow
Hydrol., 187共1–2兲, 81–104. dynamics and real network structures on Hortonian scaling of peak
Gupta, V. K., and Dawdy, D. 共1995兲. “Physical interpretation of regional flows, hydrofactals.” J. Hydrol., 322, 155–167.
variations in the scaling exponents in flood quantiles.” Hydrolog. Pro- Ogden, F. L., and Dawdy, D. R. 共2003兲. “Peak discharge scaling in small
cess., 9共3/4兲, 347–361. Hortonian watershed.” J. Hydrol. Eng., 8共2兲, 64–73.
Rodriguez-Iturbe, I., Mejia, J. M., and Dawdy, D. R. 共1972兲. “Streamflow
Downloaded from ascelibrary.org by UNIVERSITY OF VIRGINIA on 11/03/13. Copyright ASCE. For personal use only; all rights reserved.

Gupta, V. K., Mesa, O. J., and Dawdy, D. 共1994兲. “Multiscaling theory of


floods: Regional quantile analysis.” Water Resour. Res., 30共12兲, simulation: 1. A new look at Markovian models, fractional Gaussian
3405–3421. noise, and crossing theory.” Water Resour. Res., 8共4兲, 931–941.
Gupta, V. K., and Waymire, E. 共1990兲. “Multiscaling properties of spatial Water Science and Technology Board 共WSTB兲 Committee on Opportu-
rainfall and river flow distributions,” J. Geophys. Res., 95共D3兲, 1999– nities in the Hydrologic Sciences. 共1991兲. Opportunities in the hydro-
2009. logic sciences, National Research Council, Washington, D.C.

JOURNAL OF HYDROLOGIC ENGINEERING © ASCE / JANUARY/FEBRUARY 2007 / 3

J. Hydrol. Eng. 2007.12:1-3.

Вам также может понравиться