What a '100-year flood' really means

A river floods badly, and within days an official or a newspaper calls it a "100-year flood". Someone asks the obvious question: does that mean we are safe for the next hundred years? It does not. The 100-year flood meaning is statistical, not a timetable. It describes a flood size that has a 1 in 100 chance of being equalled or exceeded in any single year, and that chance does not change because a big flood happened last year. This article explains what a return period is, how the risk adds up over a lifetime, how these floods are estimated from a short record, and where the estimates are weakest.

What does "100-year flood" actually mean?

A 100-year flood is the flood peak that has a 1 percent chance of being equalled or exceeded in any given year. The United States Geological Survey (USGS) puts it plainly: the term is misleading because it suggests a flood that happens only once in 100 years, whereas it is "a statistical designation" with a 1-in-100 chance of occurring in any year.

The Indian dam-design standard defines the idea in the same way. A flood of return period T years is "expected, on the average, to be equalled or exceeded once in T years", so the probability of equalling or exceeding it in a year is 1/T [3]. The same standard adds that you must say which flood element you mean (peak discharge, volume above a threshold and so on), because a "100-year" peak and a "100-year" volume are different floods [3].

Because of the confusion, the USGS guidelines for flood frequency use the term annual exceedance probability (AEP), the probability in any year that a flood threshold is exceeded. A 100-year flood is a flood with an AEP of 0.01, or 1 percent [1].

How do you work out the return period?

The return period (also called recurrence interval) is the reciprocal of the annual exceedance probability:

T = 1 / p          p = annual exceedance probability

p = 0.10   ->  T = 10 years
p = 0.02   ->  T = 50 years
p = 0.01   ->  T = 100 years
p = 0.002  ->  T = 500 years

"Return period" is a long-run average. Floods come irregularly, and two 100-year floods can arrive in successive years, or none for 250 years, without anything being wrong with the estimate [2].

What is the chance of a 100-year flood in your lifetime?

The chance builds up with every year you stay exposed. If the AEP is p and years are independent, the probability of at least one exceedance in n years is:

P(n) = 1 - (1 - p)^n

The USGS guidelines give this relation and note that a property with a 1 percent annual chance of flooding has a 26 percent probability of being flooded over the life of a 30-year mortgage, about one in four [1]. The same arithmetic for other horizons and other flood sizes:

Check for 50 years at p = 0.01: 1 − 0.99^50 = 1 − 0.605 = 0.395, or 39.5 %. Over 100 years the answer is 63 %, not 100 %.

Why do engineers need return periods at all?

Every structure trades the cost of building bigger against the damage if it is overtopped, and return periods give that trade-off a common language. In India, the dam-design standard IS 11223 sets the inflow design flood for a small dam at the 100-year flood, for an intermediate dam at the standard project flood and for a large dam at the probable maximum flood. It allows these to be raised or lowered where the hazard downstream is unusually high or low [3]. The same standard suggests, in general, a 25-year flood for land acquisition and a 50-year flood for built-up property when working out upstream submergence [3].

The same logic sets the mapped extent in flood hazard maps and floodplain zoning; for rainfall, the equivalent is the intensity-duration-frequency (IDF) curve.

How is a 100-year flood calculated?

A 100-year flood is calculated by fitting a probability distribution to the yearly maximum flood peaks recorded at a gauging site and reading off the flow at AEP 0.01. This is flood frequency analysis, and the usual steps are these.

1. Build the annual maximum series. Take the largest instantaneous peak discharge in each year of record. The USGS guidelines recommend this series for floods with an AEP of 0.10 or smaller, and note there is little difference from a partial-duration series for such rare floods [1].
2. Check the data. The method assumes the peaks are independent, identically distributed and drawn from a stable (stationary) process. Exploratory checks for trends, shifts, mixed flood causes and measurement error come first [1].
3. Fit a distribution. The US federal guidelines (Bulletin 17C) fit a log-Pearson Type III distribution to the logarithms of the peaks [1]. Other countries and Indian studies use Gumbel, generalised extreme value or generalised logistic distributions.
4. Read off the quantile and its uncertainty. The flood for the chosen AEP is read from the fitted curve, together with confidence intervals.

A worked example (illustrative)

Suppose a gauging site has 30 years of annual peaks with a mean of 400 m³/s and a standard deviation of 150 m³/s. Fitting a Gumbel (extreme value type I) distribution, the flood for return period T is the mean plus a frequency factor K times the standard deviation, the general form set out by Chow [7]:

K = -(√6 / π) × [ 0.5772 + ln( ln( T / (T - 1) ) ) ]
Q_T = mean + K × standard deviation

T = 100:  ln(ln(100/99)) = ln(0.01005) = -4.600
          K = -0.7797 × (0.5772 - 4.600) = 3.137
          Q_100 = 400 + 3.137 × 150 = 871 m³/s

T = 10:   K = 1.305,   Q_10 = 596 m³/s
T = 50:   K = 2.592,   Q_50 = 789 m³/s

The 100-year flood here is about 2.2 times the average annual peak. In regional work, Indian agencies express the same idea as a growth factor, the ratio of the T-year flood to the mean annual peak flood (MAF). For the Middle Ganga Plains subzone 1(f), the National Institute of Hydrology lists growth factors of 0.906 at 2 years, 2.527 at 50 years and 2.840 at 100 years, so a gauged catchment with a MAF of 500 m³/s would have a 100-year flood of about 500 × 2.840 = 1,420 m³/s (the MAF is illustrative; the factors are from the published table) [5]. The Central Water Commission (CWC) publishes regional flood estimation reports for subzones of the country, including one for the Western Himalayas (zone 7) [4].

Why is the 100-year flood so uncertain?

The 100-year flood is a prediction about something that has probably never been observed at the site. A gauging record of 30 or 40 years is being stretched to describe a flood that should occur once in 100. Several things limit how far that stretch can be trusted.

• Short records. The USGS guidelines say flood records are limited and that an estimate "may be outdated a few years after it is made". New data alone can justify a fresh assessment [1]. The longer the extrapolation beyond the record, the wider the confidence interval [1].
• Measurement error at the top end. The largest peaks are the hardest to measure, and are often estimated by extending a rating curve beyond the highest direct gauging. See rating curves and stage-discharge uncertainty and measurement uncertainty for water data.
• Mixed causes of flooding. Floods in some watersheds come from different processes, such as rainstorms and snowmelt, and a record that mixes them may not be homogeneous [1]. See snow and glacier melt for Himalayan rivers.
• Skewed distributions. A few extreme years dominate the high quantiles. Bulletin 17C therefore combines the station's own skew with regional information and uses the Expected Moments Algorithm to bring in historical and paleoflood information [1].
• A changing climate and catchment. The method assumes the flood-generating process is stationary. The USGS guidelines point out that stationarity is a property of the underlying process rather than of the observed data, and that non-stationary behaviour is hard to detect in peak-flow series [1]. In our review of non-stationary flood frequency methods, we found that the field has moved towards models whose parameters vary with time or with climate drivers, such as GAMLSS, Bayesian methods and peaks-over-threshold approaches, and concluded that non-stationarity needs to be built into flood risk frameworks under future climate change [6].

So treat "the 100-year flood is 871 m³/s" as a central estimate with a range around it.

Does a 100-year flood in the Himalaya mean the same thing?

The meaning is the same; the estimate is harder. Hill catchments are steep, small and fast, and gauged records are often short, so the 100-year value rests on very little data. Where a site has no record, regional methods such as the CWC and NIH formulae are used instead [4][5]. Floods from cloudbursts, debris flows or landslide-dam breaches may be poorly represented in a few decades of gauge data. A return period computed from a gauge is a statement about the gauge's record, not a guarantee about the next monsoon. See river monitoring in hill and Himalayan catchments.

How should the term be used and reported?

• Prefer "1 percent annual exceedance probability (AEP) flood" or "a flood with a 1-in-100 chance each year" in public communication; the USGS flood guidelines use the AEP wording for this reason [1].
• Name the flood element (peak discharge, volume or level) and the site, as the Indian standard requires [3].
• State the record length and the method, because a 100-year estimate from 15 years of data is not the same kind of number as one from 80 years.
• Do not treat the design flood as a safety guarantee. Warning systems exist because floods larger than the design value, and floods that arrive in the "wrong" year, still happen. See warning level vs danger level.

Key points

• A 100-year flood has a 1 percent chance of being equalled or exceeded in any year. It is not a flood that comes once per century.
• Return period T is 1/p. The chance of at least one exceedance in n years is 1 − (1 − p)^n: 26 % over 30 years and 63 % over 100 years for a 1 percent flood.
• Flood frequency analysis fits a distribution to annual maximum peaks; the 100-year value is an extrapolation with real uncertainty.
• Short records, rating-curve error at high flows, mixed flood causes and a changing climate all widen that uncertainty.
• Return periods guide design, but they do not replace real-time monitoring and warning.

Where ClimMaTech fits

Return periods describe what a river can do over decades; a warning system has to deal with what it is doing now. At ClimMaTech we work on the second part, and the first also depends on it: long, well-checked records of water level and discharge are what make future flood frequency estimates narrower. PRAVAAH, combining an AWLR (Automatic Water Level Recorder) with an SVR (Surface Velocity Radar), records level and surface velocity at a river section, and our flood warning solution describes how such readings feed alerts when a river is rising.

References

1. J. F. England Jr., T. A. Cohn, B. A. Faber, J. R. Stedinger, W. O. Thomas Jr., A. G. Veilleux, J. E. Kiang, R. R. Mason Jr., Guidelines for determining flood flow frequency: Bulletin 17C, U.S. Geological Survey Techniques and Methods 4-B5, version 1.1, 2019 (first posted 2018). doi:10.3133/tm4B5 (full PDF: pubs.usgs.gov)
2. K. Dinicola, The "100-Year Flood", U.S. Geological Survey Fact Sheet 229-96, 1996. pubs.usgs.gov
3. Bureau of Indian Standards, IS 11223:1985, Guidelines for fixing spillway capacity, reaffirmed 2004. law.resource.org (archived copy)
4. Central Water Commission, Regional Flood Estimation Reports (list of subzone reports including Western Himalayas zone 7). cwc.gov.in
5. National Institute of Hydrology, Roorkee, Implementable Technologies for Water Resources, section "Assessment of floods in gauged and ungauged catchments". nihroorkee.gov.in
6. S. Barbhuiya, M. Ramadas, S. S. Biswal, Nonstationary Flood Frequency Analysis: Review of Methods and Models, in River, Sediment and Hydrological Extremes: Causes, Impacts and Management, Springer, 2023, pp. 271-288. doi:10.1007/978-981-99-4811-6_15
7. V. T. Chow, A general formula for hydrologic frequency analysis, Transactions, American Geophysical Union 32(2), 231-237, 1951. doi:10.1029/TR032i002p00231