, [1]
Presentation given at: III Jornades de Meteorologia Eduard Fontsere, sponsored by the Catalan Meteorological Society, 15-16 November 1997, Barcelona, Spain. The viewgraphs from this presentation can be viewed here.
Flash floods are defined to be flood events where the rising water occurs during or a matter of a few hours after the associated rainfall. If the damaging water level increases occur more than a few hours after the rainfall, the event is considered to be a flood, not a flash flood. They are a worldwide hazard, at least outside the immediate polar regions. Whereas many weather phenomena have specific geographical locations where they occur, rainfall is an event that occurs virtually everywhere. If rainfall is possible, there are going to be occasions when it becomes intense and that intensity is maintained long enough to create the potential for flash floods. Hydrology plays a large role in the flash flood problem; a given amount of rainfall in a given time may or may not result in a flash flood, owing to such factors as antecedent precipitation, soil permeability, terrain gradients, and so on. Therefore, flash flood forecasting involves both a hydrological and a meteorological forecast.
The urgency for development of reliable and accurate flash flood forecasts is growing, for a number of reasons. Simple population growth increases the number of citizens at risk, of course. Increasing urbanization alters the hydrological and mesoscale meteorological setting; cities have greater precipitation run-off than rural areas and the cities themselves have been implicated in altering the mesoscale meteorology in their vicinity (see Chagnon 1978). Greater recreational use of mountainous regions puts more people at risk (as illustrated by the recent Biescas flash flood case; see Riosalido 1997). Urban expansion is creating pressure to develop housing and commercial use in flood-prone regions, often without inhabitants being aware of the risks they are taking. Even as our scientific capabilities to forecast flash floods are increasing, the increasing hazards posed by expanding populations are resulting in a steady loss of lives and a growing financial loss associated with flash floods
In this presentation, I will review the flash flood forecasting problem mostly from a meteorological viewpoint. In section 2, I describe briefly the meteorological situations associated with flash-flood producing rainfalls. Section 3 considers progress in objective methods to forecast precipitation quantitatively, while section 4 discusses some perspectives on subjective forecasting methods. I conclude with section 5, emphasizing the need for an expanded collaboration between meteorologists and hydrologists if the problem is going to be addressed properly.
As described in Doswell et al. (1996), flash floods are the result of high precipitation rates persisting for a relatively long time (order of a few hours). In view of all the variables, especially the hydrological ones, it is not worthwhile to specify threshold values for rate and duration. The minimum amount of rainfall and its duration to create flash flood conditions depends on the hydrological setting for the event. Although it is not possible to set thresholds for rate and duration, experience makes it clear that the majority of flash flood-producing rainfall is convective. This is because the rapid upward vertical motion in convection also promotes high precipitation rates. Since not all convective storms produce flash floods, a major limiting factor is the duration of the relatively intense convective rainfalls. Most convective events do not persist in any given catchment long enough to produce flooding, so the duration of convection is the key issue. In the flash flood cases, however, the convection is maintained over a specific location; Chappell (1986) refers to these as "quasistationary convective events." What makes the convection remain geographically fixed for an extended period is the preferential development of new convective cells so as to nearly cancel the tendency of such cells to drift with the wind. Anticipating properly this "propagation" of the system is essential to making a good forecast.
Many convective systems are organized such that the individual cells within the system are aligned laterally. As described in Doswell et al. (1996), if the system motion is nearly parallel with that linear structure but the cells move roughly parallel to the line, then any given point along that line experiences several convective cells in sequence. This is the so-called "train effect" wherein several convective cells pass over a given point, more or less like railroad cars in a train. Most convectively-associated flash flood events are of this character.
Occasionally, it also turns out that supercell storms produce prodigious local rainfalls (Doswell 1994). Although supercell storms can be quite inefficient at producing rainfall from their water vapor (due to entrainment of dry air), their very strong, large, and persistent updrafts can more than compensate for any inefficiency. Supercells are known to move substantially slower than the mean winds, owing to propagation effects, so that at times they move slowly enough to become a flash flood threat. Supercell-associated flash floods are especially dangerous, as the severe weather produced by such storms can divert attention from their flash flood potential. Not all supercells are heavy precipitation producers (see Doswell and Burgess 1993), but their contribution to flash flood potential should not be taken lightly.
At times, heavy rainfalls can result not from convection but from the forced ascent of moist air in stable stratification. This is virtually always associated with orographically-forced rainfall events. Examples of this are cited in Doswell et al. (1996) and Romero et al. (1998). In some cases (like the Big Thompson event in 1972 and the Rapid City event in 1972; see Maddox et al. 1978), quasistationary convection occurs in situations involving an orographic component. It appears that in many cases associated with orographic ascent, a convective component can develop simply because ascent tends to reduce the static stability, but convection still may not be the dominant contributor to the rainfall (see Romero et al. 1998).
Progress along several distinct lines is making objective, quantitative precipitation forecasting (QPF) increasingly accurate. The development of sophisticated parameterization of convection within mesoscale numerical simulation models is offering a significant improvement in the model-generated QPF (see Sénési et al. 1996). It is now possible to consider a future when convection is treated explicitly in an operational model with grid spacings on the order of one km. A major challenge to numerical simulation models, whether they parameterize convection or not, is going to be specification of the initial conditions. Stensrud and Fritsch (1994) have shown that in their mesoscale model with parameterized convection, an accurate prediction can depend strongly on the presence or absence of mesoscale details in the initial conditions; notably, this dependency is greatest in conditions wherein the synoptic-scale processes are not very intense. Romero et al. (1998) have suggested, similarly, that the greater the convective component associated a heavy rainfall event, the less likely a mesoscale numerical simulation model is going to be successful in predicting the rainfall. Although high-resolution nonhydrostatic models can forecasts convection explicitly, it is not at all clear that they can handle the details well enough to make precise QPFs directly (see also the discussion in Brooks et al. 1992).
Advances in QPFs direct from the simulations is not the end of the progress in objective QPFs, however. It is well-known that various sorts of post-processing of model output can improve substantially upon the model output. This is exemplified by the well-known use of model output statistics (MOS; see Glahn and Lowry 1972). The technique exploits the value of statistical methods in developing QPFs using more than just the model's explicit forecasts of precipitation. Results indicate that MOS post-processing of model output can improve significantly on the model's direct QPFs.
Other model output post-processing methods besides MOS also are being explored; notably, neural networks (Hertz and Palmer 1991) and ensemble methods (Leith 1974). Much of this work is so new that it has not yet been published, but it is quite clear that they offer prospects of significant improvements over MOS. A simple PC-based neural network approach (T. Hall, personal communication) has been developed in the Fort Worth, Texas office of the National Weather Service using gridded model output as predictors and is being tested in real time. Some early results are shown in Table 1.
Table 1. Verification of Neural Net QPF (PoP). For all forecasts of area-averaged precipitation having a probability of precipitation (PoP) >= 45%, the forecast categories are compared against the observed area-averaged precipitation (measured in inches) in the Dallas-Fort Worth metropolitan area. The diagonal elements are italicized, and the number in the lower right corner is the percentage of the total number of such forecasts along the diagonal, corresponding to exactly correct category forecasts. All empty boxes correspond to zeroes. The forecasts are for 24-h precipitation and cover approximately one year's worth of forecasts. N(f) is the number of forecasts in each category and N(x) is the number of observations in each category.
O B S E R V E D
0 .01 .05 .11 .20 .30 .40 .50 .60 .70 .80 1.0 2.0
-.05 -.11 -.20 -.30 -.40 -.50 -.60 -.70 -.80 -1.0 -2.0 -3.2 N(f)
0 0
.01-.05 4 1 5
.05-.10 3 2 1 1 7
F .11-.20 2 1 7 1 1 12
O .20-.30 4 2 2 1 9
R .30-.40 5 1 4 10
E .40-.50 1 2 4 2 1 1 11
C .50-.60 3 3
A .60-.70 1 1 2 2 1 7
S .70-.80 1 2 2 5
T .80-1.0 1 1
1.0-2.0 2 3 1 6
2.0-3.2 1 1
N(x) 0 6 5 19 6 9 9 6 4 2 5 4 2 40%
Ensemble methods are also being explored. Ensemble forecasting involves several runs of a reduced-resolution version of a forecast model. These multiple runs can explore the sensitivity to initial condition uncertainty (through some objective method of perturbing the initial conditions) or to model uncertainty (through altering some internal aspect of the model during different runs). Owing to nonlinearity, it is well-known that model solutions will eventually diverge; information is contained in the spread of solutions. There is some belief that the spread within the ensemble is related to the accuracy of the ensemble mean, but this remains to be demonstrated. As it stands, the simple ensemble mean typically produces forecast errors at least comparable to those from a single, high-resolution run. Recently, some simple neural network post-processing of the ensembles shows that the QPFs associated with that post-processed information results in QPFs that are at least as good as MOS.
It is important to understand that objective forecasting is moving rapidly toward being able to express forecasts, including QPFs, in probabilistic terms. All forecasts are uncertain, and it is becoming clear that users of weather information are interested not only in our forecasts per se, but also need information about the uncertainty associated with each forecast. Thus, probabilistic approaches are virtually certain to be used more widely; probability is the language of uncertainty (Sanders 1963).
In the paper by Doswell et al. (1996), it was suggested that an ingredients-based methodology be followed in subjective forecasting of flash floods. It is useful compare the "ingredients" approach to various methods that have preceded it.
Naturally, the simplest approach is simply to take whatever objective guidance might be available and use it without any modification whatsoever. This is the so-called "meteorological cancer" method (first described by Snellman 1977). In effect, the guidance becomes the forecast; for my purposes, this scheme is useful primarily as a baseline for comparison . Clearly, when human intervention provides no accuracy improvement over objective guidance, this suggests that whatever methods humans are using to make changes to the guidance are not particularly beneficial. Note that any assessment of forecasts must begin with a careful and comprehensive verification of those forecasts; any forecast that cannot be verified or that is not verified (for whatever reason) offers no opportunity for forecast improvement. Thus, I am going to assume that both objective guidance and subjective forecasts are being verified as rigorously as it is possible to do. Verification is an important topic in its own right, but will not be discussed here.
What are some methods forecasters have used in the past to make forecasts? Unfortunately, there has been virtually no systematic study of human forecast methods, in spite of the growth in such fields as cognitive psychology. To date, we have precious little information about how humans go about making forecasts. I see this as a major barrier to making progress in subjective forecasting, but I can do little to rectify this deficiency at present. Thus, what I have to present is virtually entirely anecdotal.
Subjective forecasting methods include: pattern recognition, checklists of parameters, analogs, empirical "rules of thumb," and so forth. Many of these involve sets of "ingredients" that may or may not have a clear physical relationship to the events in question. Most subjective methods are highly personal, although some may have evolved under the tutelage of an experienced forecaster. Typically, the methods have not been subjected to a rigorous verification over a wide sample of events. Often forecasters who try a new approach will abandon it quickly if it fails in its first few applications; conversely, they will continue to use it if that method provides them with immediate success (even if it subsequently fails them). Thus, subjective forecasting is prone to many sorts of human failings, and it often has little respect among non-forecasting scientists.
On the other hand, at its best, subjective forecasting involves a complex integration of quantitative and qualitative factors that provide seemingly impossibly good forecasts. Human forecasters are capable of forecasting success well beyond what objective methods provide. By processes which are not even vaguely understood, humans seem able, at times, to use their experience and knowledge to make significant improvement over objective guidance products, irrespective of the source of the guidance.
Objective guidance has great difficulty with rare events, so the rare events offer humans considerable opportunity to improve upon that guidance. In many situations involving flash floods, the events are rare from a meteorological and hydrological viewpoint. Hence, it is plausible to suggest that humans ought to be an important component in accurate forecasts of flash floods. It has been my experience, however, that many forecasters are not up to this challenge; flash flood events have a pretty poor forecasting record associated with them. I have no interest in establishing blame for this. Nevertheless, I believe that many flash flood events (but by no means all such events!) could have been anticipated. Only a part of the reasons for this lie in the methods used by forecasters.
I believe that forecasters need to consider the range of possibilities inherent in every day's weather patterns. Are the ingredients for a particularly hazardous type of weather event (e.g., a flash flood) present or nearby? Part of the answer for getting forecasters to recognize the threats in a day's weather is education and training. I will not dwell on this, as I have written extensively on it elsewhere (e.g., Doswell et al. 1981; Doswell 1986a).
Another part of the answer is for forecasters to be more involved in diagnosis of the data (see Doswell and Maddox 1986). A forecast can be thought of as the combination of a diagnosis with a trend (Doswell 1986b). If Q is some forecast quantity, then, a mathematical statement of the forecast problem is contained in the simple Taylor's series expansion of Q:
, [1]
where the analysis is identified with Q at time to and the trend at that time is identified with dQ/dt. Doing an analysis of the Q-field is not sufficent, however, for a diagnosis. A diagnosis involves a multivariate, multidimensional interpretation of all the data designed to explain why the Q-field has the particular distribution it has at time to. Often, the trend needs to become nonlinear (i.e., the higher order terms in [1] need to be considered). Purely linear extrapolation has very limited value, especially in the context of forecasting convection, where time scales are on the order of 10s of minutes.
In today's technology-dominated forecast offices, I see forecasters inundated with numerical model output and a host of tools designed to manipulate it. On the other hand, the observational data and its diagnostic tools have not been given anywhere near the attention they deserve. Hence, forecasters spend a lot of time trying to resolve what is typically conflicting guidance products. My argument is that this is mostly a waste of precious time. If forecasters do not have the capability to explain why the weather is in its present state, how can we reasonably expect them to make a forecast (that is more useful than the objective guidance) of some future state of the weather?
The idea of ingredients-based forecasting is to reduce the problem of forecasting a particular event to its simplest possible terms. What is essential for having an event of that type? If it is already under way, is it going to continue? If it is not underway at present, will its ingredients be concatenated at some future time?
Space simply does not permit an extensive discussion. Obviously, objective QPF methods are going to improve in the future as models and the methods for processing the model output improve. Hopefully, there eventually will be a growth in the recognition for improved forecaster education and training to improve upon the objective methods. I have been somewhat frustrated to see how slowly an integration of the meteorological and hydrological components of the flash flood forecasting system has evolved. There have been some limited experiments, but we are a long way from having a true operational hydrometeorological forecasting system appropriate for flash floods. Some of the lack of interaction stems from the respective communities having different foci: hydrologists seem to be primarily interested in detecting rainfall after it has commenced (i.e., as input to the hydrological prediction models), whereas meteorologists seem mostly interested in forecasting rainfall before it begins.
Given the uncertainties in forecasting rainfall, hydrologists might well be skeptical of a meteorologist's rainfall predictions; I believe that eventually the hydrologists will be more willing to incorporate rainfall forecasts if we meteorologists can re-cast our operational QPFs in probabilistic terms (see Krzysztofowicz et al. 1993). When meteorological and hydrological models are married so as to provide objective guidance relative to true flash flood potential (rather than just heavy precipitation potential), then human forecasters will be confronted with the challenge to use this information wisely in producing the best possible forecasts to the forecast users. It is important to remember that a forecast convective weather event that is not anticipated is not likely to be detected rapidly enough to do the users any good. Flash floods almost certainly will be low-probability events (especially at a given location), right up to the time that rainfall actually begins. The key to good flash flood warnings (associated with detecting the event as soons as possible after it begins) is having recognized that a flash flood event was possible (i.e., an enhanced probability over climatology) before the rainfall ever began. There is little doubt that this is a reasonable expectation from a forecasting system that includes both high-quality forecasting guidance and human forecasters who are educated and trained to improve upon that guidance.
Acknowledgments. I appreciate the input provided to me by Drs. Harold Brooks and Steve Mullen.
Brooks, H.E., C.A. Doswell III and R.A. Maddox, 1992: On the use of mesoscale and cloud-scale models in operational forecasting. Wea. Forecasting, 7, 120-132
Chagnon, S.A., 1978: Urban effects on severe local storms at St. Louis. J. Appl. Meteor., 17, 578-586.
Doswell, C.A. III, 1986a: The human element in weather forecasting. Nat. Wea. Dig., 11, 6-18.
______, 1986b: Short range forecasting. Mesoscale Meteorology and Forecasting, Chapter 29 (Peter Ray, Ed.), Amer. Meteor. Soc., 689-719.
______, 1994: Flash flood-producing convective storms: Current understanding and research. Report of the proceedings, U.S.-Spain Workshop on Natural Hazards (Barcelona, Spain), National Science Foundation, 97-107.
______, and R.A. Maddox, 1986: The role of diagnosis in weather forecasting. Preprints, 11th Conf. Wea. Forecasting and Analysis, Kansas City, MO, Amer. Meteor. Soc., 177-182.
______, L.R. Lemon and R.A. Maddox, 1981: Forecaster training - A review and analysis. Bull. Amer. Meteor. Soc., 61, 983-988.
______, H.E. Brooks and R.A. Maddox, 1996: Flash flood forecasting: An ingredients-based methodology. Wea. Forecasting, 11, 560-581.
______, C. Ramis, R. Romero and S. Alonso, 1998: A diagnostic study of three heavy precipitation episodes in the western Mediterranean. Wea. Forecasting, [in press].
Glahn, H.R., and D.A. Lowry, 1972: The use of model output statistics in objective weather forecasting. J. Appl. Meteor., 11, 1203-1211.
Hertz, J., A. Krogh and R.G. Palmer, 1991: Introduction to the Theory of Neural Computation. Addison-Wesley, 414 pp.
Krzysztofowicz, R., W.J. Drzal, T.R. Drake, J.C. Weyman and L.A. Giordano, 1993: Probabilistic quantitative precipitation forecasts for river basins. Wea. Forecasting, 8, 424-439.
Leith, C.E., 1974: Theoretical skill of Monte Carlo forecasts. Mon. Wea. Rev., 102, 409-418.
Maddox, R.A., L.R. Hoxit, C.F. Chappell and F. Caracena, 1978: Comparison of meteorological aspects of the Big Thompson and Rapid City flash floods. Mon. Wea. Rev., 106, 375-389.
Riosalido, R., 1997: Estudio meteorologico de la situacion del 7 de agosto de 1996 (Biescas). Instituto Nacional de Meteorologia.
Romero, R., C. Ramis, S. Alonso, C.A. Doswell III and D.J. Stensrud, 1997: Mesoscale model simulations of three heavy precipitation events in the western Mediterranean region. Mon. Wea. Rev., [in press].
Sanders, F., 1963: On subjective probability forecasting. J. Appl. Meteor., 2, 191-201.
Sénési, S., P. Bougeault, J.-L. Chèze, P. Cosentino, and R.-M. Thepenier, 1996: The Vaison-La-Romaine flash flood: Mesoscale analysis and predictability issues. Wea. Forecasting, 11, 417-442.
Snellman, L.W., 1977: Operational forecasting using automated guidance. Bull. Amer. Meteor. Soc., 58, 1036-1044.
Stensrud, D.J., and J.M. Fritsch, 1994: Mesoscale convective systems in weakly forced large-scale environments. Part II: Generation of a mesoscale initial condition. Mon. Wea. Rev., 122, 2068-2083.