Frequency and duration of floods are analyzed using the global flood database of the Dartmouth Flood Observatory (DFO) to explore evidence of trends during 1985–2015 at global and latitudinal scales. Three classes of flood duration (i.e., short: 1–7, moderate: 8–20, and long: 21 days and above) are also considered for this analysis. The nonparametric Mann–Kendall trend analysis is used to evaluate three hypotheses addressing potential monotonic trends in the frequency of flood, moments of duration, and frequency of specific flood duration types. We also evaluated if trends could be related to large-scale atmospheric teleconnections using a generalized linear model framework. Results show that flood frequency and the tails of the flood duration (long duration) have increased at both the global and the latitudinal scales. In the tropics, floods have increased 4-fold since the 2000s. This increase is 2.5-fold in the north midlatitudes. However, much of the trend in frequency and duration of the floods can be placed within the long-term climate variability context since the Atlantic Multidecadal Oscillation, North Atlantic Oscillation, and Pacific Decadal Oscillation were the main atmospheric teleconnections explaining this trend. There is no monotonic trend in the frequency of short-duration floods across all the global and latitudinal scales. There is a significant increasing trend in the annual median of flood durations globally and each latitudinal belt, and this trend is not related to these teleconnections. While the DFO data come with a certain level of epistemic uncertainty due to imprecision in the estimation of floods, overall, the analysis provides insights for understanding the frequency and persistence in hydrologic extremes and how they relate to changes in the climate, organization of global and local dynamical systems, and country-scale socioeconomic factors.
Higher levels of vulnerabilities to extreme events, especially floods, are
becoming a “new normal” in both developing and developed countries
Other impacts of floods include various deteriorations of social services,
economic disruptions, health-related issues, and consequences of population
displacement (i.e., disturbances in the food supply chain, undernutrition,
water-/vector-borne diseases, and being injured, displaced, or left homeless)
Often, these impacts are magnified when the floods are due to persistent and
recurrent rainfall. Such floods typically last longer (henceforth called long-duration floods) and are associated with repeated rainfall events in the
regions. Recently,
Global and near-daily observations from the Earth's surface are now
available through satellite microwave sensors (active/passive), which are
being employed to measure the changes of water surfaces (e.g., river
discharge and watershed runoff)
Given the floods (especially the long-duration floods) are caused by a
systematic organization of the global-to-local dynamical systems of climate
and atmosphere
Consequently, we utilized the global active archive of flood events (with
31 years of data from 1985 to 2015) to address the following five questions:
How has the annual frequency of floods changed at the global scale and
various latitudinal belts during the last 3 decades? How has the probability distribution of flood duration (represented by the moments
and extreme values) changed at the global scale and various latitudinal belts
during the last 3 decades? Are the changes (if any) in the flood frequency and the probability
distribution of flood durations due to the changes in a specific flood class,
i.e., short, moderate, or long duration? Can the changes (if any) in the flood frequency and the probability
distribution of flood durations be related to the variability in the atmospheric
teleconnections and low-frequency climate oscillations? Which countries are most vulnerable to short-, moderate-, and long-duration floods?
We address each question using a formal hypothesis-testing framework. This
paper is organized as follows: Sect. 2 provides the detailed information
about the global flood database, design hypotheses, and employed methodology
in this study. Section 3 presents the results of the hypothesis tests and the
country-scale vulnerability analysis to different flood durations. In Sect. 4,
we present a generalized linear model (GLM) framework to investigate
the potential causes of the observed trends and also discuss the other
comparable global trend studies. Finally, we present the concluding remarks
and highlights in Sect. 5.
A comprehensive record of flood events is available from the Dartmouth Flood
Observatory (DFO) founded in 1993 at Dartmouth College, NH, United States.
In 2010, the observatory moved to the Community Surface Dynamics
Modeling System (CSDMS) (
It is important to note that the quality of data has improved in
recent
times. The improvements in the level of media reporting and information
quality have improved the reliability of the data. At the same time, the
likely improvements in the accuracy of in situ measurements, advances in
satellite and ground-based sensors, data storage, and transfer facilities
also contributed to the data quality. Moreover,
The flood events are spatially aggregated to five climate zones – tropics
(23.5
Spatial segmentation to assign the global flood events (1985 to
2015) into different latitudinal belts: midlatitudes (N): 35–55
Next, for each latitudinal belt, the total number of floods per year
(calendar year from 1 January to 31 December), the duration of these floods,
and their location (name of country) are processed. This procedure is
formulated as follows:
In addition, the number of floods in each latitudinal belt are also
categorized in terms of their duration. We denote the event as a
short-duration flood
We used large-scale ocean–atmospheric teleconnections to investigate the
extent to which the trends in the floods can be related to natural
variability
We obtained 31 years (1985–2015) of ENSO data (aggregated based on the
monthly anomalies of Niño 3.4) from the HadISST1 dataset
In addition to the frequency of the floods (
Proposed hypotheses and evaluation approach.
We calculate the MAD of flood duration as an
indicator of the deviation from the central tendency. The MAD is a robust
measure to quantify the within-year variation in flood duration. It is a good
measure of scale for distributions with heavier tails
The presence of outliers amongst the variables will generate a large and
possibly misleading measure of skewness
Note that the sample sizes (number of floods) may be different for different years. For instance, the total number of floods in 1985 at the global scale is 69. We compute the median, MAD, skewness, and the 90th percentile of the duration for these 69 events. Similarly, the total number of floods in 2015 at the global scale is 101, and we compute the median, MAD, skewness, and the 90th percentile for these 101 events. After obtaining the time series of these metrics, we then investigate for monotonic time trends.
For a specific country, we calculate the relative flood frequency of short, moderate, and long durations with respect to the total flood events occurring in that country. This can help us identify what flood duration class has occurred more frequently from 1985 to 2015 in that country. Correspondingly, the reported flood damage for that event has also been noted along with its relative damage in reference to the total flood damages in that country from 1985 to 2015.
In order to investigate the association between flood duration and damage at
the country scale, we present a linear model for flood damage (
Most of the global precipitation studies indicate that there is a recent
increase in both the annual precipitation and extreme rainfall intensities
We begin our investigation with H1, the hypothesis that there is no
monotonic trend in the annual frequency of the flood events. We test this
hypothesis using the Mann–Kendall (MK) trend test
In hypothesis H2, we explore whether there is a change in the probability distribution of the flood duration over time. We test this hypothesis by applying the MK trend test on the three resistance moments (median, MAD, and skewness) and the 90th percentile (extreme flood duration) of the annual distribution of the flood duration. H3 is intended to investigate the changes in the patterns of flood frequencies for each category: short-, moderate-, and long-duration floods. Lastly, in H4, we investigate the potential large-scale atmospheric teleconnections to which the observed trend(s) in H1 and H2 can be related by using a GLM framework.
Our hypothesis (H4) is that the detected time trend is due to
cyclical climate influences (i.e., oscillatory behavior) associated with the
large-scale ocean–atmospheric interactions. Hence, for all the cases in
which
the null hypothesis of no trend is rejected, we attempted to
understand whether the trend relates to large-scale climate oscillations. For
this purpose, we employed a GLM framework on the
time series of the above-developed metrics with ENSO, AMO, PDO, and NAO as
covariates. GLMs are the mathematical extension of classical linear
regression models to include a broad class of model assumptions such as
linear, Poisson, exponential, log-linear, and so on with specified link
functions
The MK test (Eqs. A1–A3) is applied to each time series of
A total of 4311 flood events occurred during last 3 decades worldwide.
The results of the MK test on the annual frequency of global floods indicate that
there is a statistically significant monotonic trend with
Summary of trend analysis (Mann–Kendall test with a significance level
Frequency of flood events at the global scale and the latitudinal
scales (i.e., tropics, subtropics – N, subtropics – S, midlatitudes – N, and
midlatitudes – S); a LOESS curve fitting is shown (solid line) for the
time series in which a significant trend in the number of flood events is observed
(Mann–Kendall test with significance level
The MK trend tests are performed on the time series of the median, MAD, resistant skewness, and the 90th percentile of the flood duration. The following four subsections elaborate the results for each metric.
Same as Table 2 but for the median of flood durations.
Same as Fig. 2 but for the median of flood durations.
From Fig. 3, we can see that there is a statistically significant monotonic trend in the median of the flood duration at the global scale and all sub-spatial scales. We see that the median of the flood duration at the global scale has increased steadily from 4 days in the year 1985 to 10 days in the year 2015, indicating that the median flood duration changed to moderate duration in 2015 from short duration in 1985. In other words, it shifted one class from being less than 1 week to between 1 week and 3 weeks. Similar shifts can be observed in the tropics and the subtropics. In Table 3, we present the statistics of the tests. As in the case of the frequency of floods, we urge caution in interpreting the trends seen in the midlatitudes (S) due to the presence of zeros.
Same as Fig. 2 but for median absolute deviation (MAD) of flood durations.
The MK trend test is performed on the MAD of flood duration (Eq. 4) at the different global and latitudinal scales and presented in Fig. 4 and Table 4.
The output statistics show that there is a significant increasing trend in MAD at the global scale, and in the tropics and subtropics (N). It is interesting to note that the MAD has essentially remained constant, around 2–3 days from 1985 to 2000, and has increased since to around 5 days in 2015, indicating increased variability in flood durations within years in these belts recently. There is no significant change in the variability in the midlatitudes (N and S) and subtropics (S).
Same as Table 2 but for the median absolute deviation (MAD) of flood durations.
Same as Fig. 2 but for the resistant skewness of flood durations.
Same as Table 2 but for the resistant skewness of flood duration distributions.
The resistant skewness of flood duration is calculated for each time series using Eq. (5) and presented in Fig. 5. As before, the MK trend test is applied to these time series. A statistically significant trend in the skewness is observed at the global scale and tropical and subtropical (S) latitudes. Similar to Tables 2–4, in Table 5 we present the test statistics. We observe that the yearly asymmetrical/symmetrical behavior of the distribution of flood durations has considerably changed during the last 3 decades (from 5 to 8 approximately), with a more significant tendency towards high skewness. This change towards a right-skewed-type distribution of flood durations (e.g., from 5 to 8) can be due to the increase in occurrence of moderate- or longer-duration floods. Conversely, there is no significant trend in the skewness of flood duration in the subtropics (N) and midlatitudes (N) at the 5 % significance level.
Finally, we test for monotonic trend in the extreme values (expressed here as
the
90th percentile) of flood duration. This measure serves as
a surrogate for extremely long-duration flood events each year. By
definition, the 90th percentile of the flood duration (
The extreme duration of floods has substantially changed over the last 3 decades at the global scale, tropics, midlatitudes (N and S), and subtropics (S), as presented in Table 6. The null hypothesis that there is no monotonic trend in the tails is rejected in all regions, except the subtropics (N). Furthermore, we find that the extreme values of the long-duration flood events are more than 30 days in the recent decade, whereas they were less than 20 days in the 1980s and 1990s. The increase was monotonic.
Same as Fig. 2 but for the 90th percentile of flood durations.
Same as Table 2 but for the 90th percentile of flood duration distributions.
The highlights of trend analyses presented in Figs. 3 to 6 and Tables 3 to 6 are outlined below:
Given that we find statistically significant trends in the tails of the
distribution (90th percentile of the duration of floods),
we were interested in exploring whether there would be a trend in the
frequency of the long-duration floods as well. To investigate this, we
performed the MK test on the frequency of long-duration floods (
Summary of trend analysis (Mann–Kendall test with a significance level
As it can be seen from Table 7, there is no monotonic trend in the frequency of short-duration floods occurring across all the spatial scales, indicating that the number of short-duration floods has not changed significantly over the last 3 decades worldwide. However, this phenomenon is not true for moderate- and long-duration floods. In fact, the frequency of both moderate- and long-duration floods has increased in the tropics. There is also an increasing trend in moderate-duration floods in the subtropics (S) and long-duration floods in the midlatitudes (N). These findings are consistent with the results from H2, where we see a trend in the skewness and the tails of floods in these belts. An increase in the frequency of moderate- and long-duration floods will result in a shift of the quantile of flood duration distribution, thereby changing the skewness and the tails.
For the long-duration flood events in the tropics, the total number of events has increased from 60 before 2000 to 249 after 2000. Similarly, the total number of events in the midlatitudes has increased from 27 to 70 post-2000. In other words, there are 4 times more long-duration floods that occurred during the most recent 15 years than before the year 2000. The increase across the midlatitudes (N) is around 2.5 times pre- and post-2000.
There were 4311 flood events that occurred from 1985 to 2015 around the
world. According to Tables 2 and 7, globally, the total number of short-,
moderate-, and long-duration flood events was 2508 (
For this purpose, we first excluded countries which had less than 31 flood events to ensure that we investigate only those counties that have experienced at least one flood per year on average. This screening resulted in 28 countries with a minimum of 31 flood events during the last 3 decades. These 28 flood-prone countries are sorted as follows: the United States (388 events), China (344 events), India (226 events), Indonesia (190 events), Philippines (181 events), Australia (121 events), Vietnam (107 events), Brazil (96 events), Bangladesh (88 events), Mexico (80 events), Iran (77 events), Afghanistan (74 events), Russia (69 events), Thailand (66 events), Pakistan (66 events), Nigeria (57 events), Malaysia (54 events), Kenya (49 events), Canada (48 events), Colombia (44 events), Peru (43 events), Turkey (41 events), Nepal (40 events), France (40 events), Romania (38 events), Ethiopia (35 events), Somalia (34 events), and New Zealand (31 events).
Then, the fraction of flood frequencies for each country and duration class – short,
moderate, and long – is calculated. Figure 7a presents these
fractions for the 28 countries using the ternary plot. For 23 of these
countries, we have the data on the damages due to the floods. We computed the
expected value of the damages for each country and plotted the fractional
damage due to short-, moderate-, and long-duration floods as the second ternary
plot in Fig. 7b. The color bars indicate the total number of events
(Fig. 7a) and the total flood damage (Fig. 7b). In each plot, the
location of the country shows the relative fraction of short-, moderate-, and
long-duration flood frequency and damage. For example, in Fig. 7a, the
United States is identified as a red circle in the top corner with
To further understand the relation between flood duration and flood damage,
we fit log-linear models given in Eq. 6 for four selected countries: the United States,
Thailand, India, and China. The results of the log-linear models for these
four countries are shown in Fig. 8a. These countries are selected because
they have the highest number of long-duration floods among all countries
(Fig. 8b). Parameter
According to the DFO flood data from 1985 to 2015, the ranking results show that the frequency of short-duration floods for the United States, China, India, and the Philippines is respectively 255, 173, 133, and 122. For moderate-duration floods, the countries of China, the United States, India, and the Philippines have experienced 118, 101, 74, and 52 flood events, respectively. The long duration floods were seen mostly in India (55 events), China (53 events), the United States (32 events), and Thailand (20 events) from 1985 to the end of 2015. It should be noted that here we only presented the top 21 countries in each category.
As discussed in this section, the consequences of floods of different durations should be paid attention to, as this plays a big role in designing appropriate flood-proofing infrastructure and developing early warning systems and flood insurance payout structures. The relation between the duration of floods and the induced damages, and how they might vary across different countries, was also investigated here.
The trends in the frequency and the distribution of the floods (prominent in long-duration floods) may be related to several causes ranging from measurement uncertainty in the DFO flood data, climate and atmospheric teleconnections, and socioeconomic contributions such as the increased exposure to the flood events. We attempt to explain these possibilities in the following two sections.
The flood archive data provided by the DFO have been collected using different
methods of observation and validation since 1985 (see the summary of the
methods in
To validate the DFO's flood statistics, we have corroborated the DFO floods
with the available in situ streamflow observations from the
GRDC (The Global Runoff Data Centre, 56068 Koblenz, Germany, 2013,
While understanding such uncertainties is essential, especially while
interpreting trends in limited data, it is also documented in the literature
that there has been an increased exposure to floods in recent times. The
number of people, residential and industrial properties, and assets exposed to
the flood events has drastically increased
While exposure of people to floods is the main concern in developing
countries, exposure of assets and properties to floods is the vital concern
for the developed countries
The frequency of heavy precipitation events has increased at the global scale
Theoretical studies also discussed the fact that mean global precipitation intensity
increased by 1–3 % (conditional on available energy budgets) in proportion
to the 1
The space–time distribution of these precipitation regimes is potentially
related to the large-scale ocean–atmosphere circulations
Hence, in an effort to investigate any significant relationship between the
observed trend in the flood data (characterized in H1 and H2) and the
variability in the climate and atmospheric circulation patterns, we
considered large-scale atmospheric teleconnections and climate indices (with
quasi-periodicity in nature that can lead to wet–dry regimes) to explain the
trend, i.e., to place the short-term trends within a longer climate
variability context as argued by
Our hypothesis (i.e., H4) is that the detected time trend is due to
cyclical climate influences (i.e., oscillatory behavior) associated with the
large-scale ocean–atmospheric interactions as recorded in the ENSO, AMO, PDO,
and NAO indices. The corresponding residual time-trend analysis from the
models explains whether the long-term natural variability dominates the
trends. We considered Poisson distribution as the link function for
ENSO, AMO, and NAO are related to We did not find any significant climate indicators that can explain the
variability in the median of the floods except for the midlatitudes (S). However,
as we pointed out before, given the limited data available at this
latitudinal belt, we do not further interpret these climate indicators as
causing the trends. There should be one or a set of inexplicable factor(s)
beyond climate teleconnections that might drive the observed trend
in AMO and NAO have an association with
Summary of generalized linear model (GLM) results relating selected
predictors to flood frequency (
To our knowledge, this study is the first analysis of global flood
events that exclusively focuses on the variability in the flood
duration using the DFO dataset over the last 3 decades (i.e.,
1985–2015). In this part, we are corroborating the presented results here
with the most relevant previous studies. A high number of recent flood
studies have focused on the regional scale, and/or have used the flood
duration to calculate the flood magnitude (i.e.,
Several flood-related studies analyzed the trends in the annual maximum
streamflow and/or precipitation across multiple spatiotemporal scales. For
example, an increasing trend in annual maximum precipitation intensities was
found by
A global assessment of flood events is performed here, focusing on the flood
frequencies and duration characteristics at different
global–latitudinal–country scales from the year 1985 to 2015. The
comprehensive assessment of the frequencies of flood events and characteristics
of the probability distribution of flood durations presented here is the very
first large-scale study of “actual” flood events worldwide focusing on
understanding the temporal changes over the last 3 decades. It was
verified here that the frequency of floods increased at the global scale,
tropics, subtropics (S), and midlatitudes (S). Selected metrics of the flood
duration showed a monotonic increasing trend for the median (at all spatial
scales), MAD (across the globe, tropics, and subtropics – N), resistant
skewness (across the globe, tropics, subtropics – S, and midlatitudes – S),
and extremes (all spatial scales except the subtropics – N). More importantly, we
find that the frequency of moderate- and long-duration floods has increased
recently, but remains unchanged for the short-duration floods at all spatial
scales. The trends in the flood frequency and extreme durations at a global
scale can be largely ascribed to ENSO, AMO, and NAO, the interannual to
decadal to multidecadal modes of variability, while the trend in the median
flood durations remains unexplained. An overall summary is presented below.
The frequency of flood events has increased; the year 2003 is recognized as
the year with the maximum number of flood occurrences across all spatial
scales; however much of this increase is within the long-term decadal to
bi-decadal climate cycles.
There is a statistically significant trend in the moments of the flood
duration at the global scale, tropics, subtropics, and midlatitudes; the
extreme floods post-2000 are more than 30 days as opposed to less than 20 days
in the 1980s and 1990s. These trends in extreme flood durations ( The yearly number of moderate- and long-duration flood occurrences increased
(from before to after the 2000s) by a factor of 4 and 2.5 events per year
across the tropics and midlatitudes (N), respectively. There was no monotonic trend observed in the frequencies of short-duration
floods (i.e., flood duration of 1 to 7 days) across all the spatial scales. Comparison of the DFO flood events with the corresponding GRDC streamflow
over the midlatitudes (N) and subtropics (N) (locations that had common records)
reveals that the reported flood events by the DFO are reasonably reliable. For
instance, 90 % of the events contain less than 7 days of deviation in their flood durations.
In addition, we also presented a simple overview of the vulnerability profile for different countries. This can be helpful to inform and improve the flood warning systems tailored to the various types and resource management practices during the post-disaster responses. Furthermore, with increasing globalization, countries are now interdependent through supply chain networks to achieve streamlined production and overall cost reductions. A country-level understanding of the exposure to different types of floods can help more accurately predict the vulnerable nodes that might cause a systemic network failure. It can also provide the necessary analysis for pricing and portfolio risk management for the agencies that insure and hedge against the flood losses.
While this study explores the trends in the frequency and duration of global floods, especially the long-duration floods, it is necessary to investigate the cause–effect mechanism of these trends along with socioeconomic variables to fully understand the emergence of floods. Understanding these hierarchical layers will provide us with comprehensive information and realization that can be translated into better defining the multiscale flood risk management and damage control strategies.
All data needed to evaluate the conclusions in the paper
are present in the paper and/or the Appendix. The data can be directly
downloaded from
The nonparametric rank-based Mann–Kendall (MK) test is widely applied to
detect the monotonic trend (i.e., a gradual change over time with consistency
in direction) in climatic or environmental time series
it is a large positive number: an upward trend is observed since the
later-measured values tend to be larger than earlier ones; it is a large negative number: a downward trend is indicated since the
later values tend to be smaller than earlier ones; it is an absolute small number: no trend is indicated.
Further, the Kendall's tau (
We validated the reported flood statistics in the DFO database with in situ
discharge observations from the Global Runoff Database from GRDC (the Global
Runoff Data Centre, 56068 Koblenz, Germany, 2013,
Summary of GRDC stations (
We employed the following procedure to validate this common record.
Three flow exceedance thresholds ( The start and end dates of a flood event in a year based on the DFO
database are delineated from the daily time series of the GRDC streamflow in that year.
Then, the total number of day(s) within the DFO's flood span when the daily
streamflow exceeds the threshold ( The difference between these two estimates is calculated as
If the GRDC flood duration is as long as the flood duration of the DFO, we
consider this to be a perfect match and the difference is 0. If GRDC did not
exhibit a threshold exceedance flow during the DFO span, we consider this to
be
a miss and the difference will be as high as the flood duration for the DFO.
Hence the absolute error is between 0 and
At the global scale and over the midlatitudes (N), for a threshold of the 90th percentile, up to 90 % of the events have an error of less than 7 days, indicating that the GRDC stations had experienced threshold exceedance floods when the DFO reported a flood. Even if we increase the threshold to the 95th percentile, we still have up to 85 % of the events with a deviation of less than 7 days. A similar pattern is seen for the subtropics (N). We refrain from interpreting the error results for the other spatial units as most of the GRDC matching data are only found in the midlatitudes (N) and subtropics (N).
Despite certain uncertainties in calculating flood duration (such as the distance between the GRDC station and the location of a flood event, anthropogenic inputs to the nature of flow rates, and a physical streamflow exceeding threshold that could precisely mimic the occurrence of a realistic flood event), it can be concluded that around 80 % of GRDC stations in this comparison could verify that the recorded flood information in the DFO including the start–end dates and flood duration parameters is reliable and would provide a certain path towards assessment of global flood events since 1985.
We corroborated the global DFO's flood frequency with the flood frequency
data available at global scale from the EM-DAT database (the Emergency Events
Database,
We investigated the DFO-reported flood events from 1985 to 2015 in terms of
the distribution of the flood beginning date and flood end date within
each month. For the starting date of flood, there are less than 5 %
(194 events) out of 4311 events that have been reported with the flood beginning
date as the middle of the month. There are 282 events reported on the first
day of the month. Together, the first and the middle day of the month account
for a total of 476 out of 4311 events (
Frequency of flood events from the DFO database and EM-DAT at the global scale (1985–2015).
Distribution of the start
Comparing flood duration (
The authors declare that they have no conflict of interest.
This article is part of the special issue “Hydro-climate dynamics, analytics and predictability”. It is not associated with a conference.
We are thankful to the Dartmouth Flood Observatory, University of Colorado at
Boulder, CO, United States, for providing the flood data. The ground-based streamflow
observations were provided by the Global Runoff Database at GRDC (The Global
Runoff Data Centre, 56068 Koblenz, Germany, 2013, Department of Energy Early CAREER award no. DE-SC0018124 for Naresh Devineni; National Science Foundation, Paleo Perspective on Climate Change (P2C2) program award no. 1401698; National Science Foundation, Water Sustainability and Climate (WSC) program award no. 1360446.
We also thank the anonymous reviewers and the editor, whose comments have
helped in improving the paper significantly.
Edited by: Julia Hall
Reviewed by: three anonymous referees