The Glomma River is the largest in Norway, with a catchment area of 154 450 km
Floods are among the most widespread natural hazards on Earth. The impacts, destruction, and costs associated with hazardous floods are increasing in concert with climate change and increase in economic value within areas susceptible to floods, a tendency most likely to strengthen in the decades to come (e.g. Alfieri et al., 2017; Hirabayashi et al., 2013; IPCC, 2012). In Europe, spatial flood patterns are changing in terms of both timing and magnitude (Blöschl et al., 2017, 2019), challenging us to examine new ways of interlinking not only different types of data, but also flood information on different timescales. Earlier studies have shown that uncertainties can be reduced if, for instance, historical data are included in estimation of floods with long return periods (e.g. Brázdil et al., 2006a; Engeland et al., 2018; Macdonald et al., 2014; Payrastre et al., 2011; Schendel and Thongwichian, 2017; Stedinger and Cohn, 1986; Viglione et al., 2013). Here we seek to extend the possibility of using historical data by including time series of reconstructed floods based on lake sediment archives which can retain imprints of past flood activity (Gilli et al., 2013; Schillereff et al., 2014; Wilhelm et al., 2018). The ultimate goals of this exercise are to (i) reduce uncertainty associated with flood prediction and (ii) provide additional insight into flood variability on longer timescales and thereby improve our understanding of how climate change impacts floods.
In many European countries, flood mitigation measures aim to reduce the
exposure and vulnerability of the society to floods. Examples of such
measures can include reservoirs, flood safe infrastructure, and land-use
planning in flood-exposed areas. These mitigation measures require estimates
of design floods, i.e. the flood size (typically given in m
Both challenges can be addressed by using data covering longer time periods, including historical data (e.g. Benson, 1950; Brázdil et al., 2006b; Macdonald et al., 2014; Schendel and Thongwichian, 2017; Viglione et al., 2013) and/or paleoflood data (e.g. Benito and O'Connor, 2013). The fact that sediment deposits can be unambiguous evidence of past floods has been documented in many studies since 1880 CE (Bretz, 1929; Dana, 1882; Tarr, 1892), and an early example of how to estimate discharge associated with giant paleofloods can be found in Baker (1973), whereas paleoflood hydrology as a concept and terminology was first introduced by Kochel and Baker (1982).
In order to include information on past floods in flood frequency analysis, it is necessary to estimate the flood sizes in m
Lakes fit for using lacustrine sediments to analyse flood frequencies are typically found where (i) flood sediments are preserved at the bottom of lakes, (ii) there is a detectable on–off signal for sediments left by floods, and (iii) there is a distinct contrast between flood deposits and regular background sedimentation (Gilli et al., 2013). Detection of flood layers in the cores can be based on X-ray fluorescence (XRF) scanning (e.g. Czymzik et al., 2013; Støren et al., 2016), magnetic susceptibility (MS) measurements (e.g. Støren et al., 2010), computed tomography (CT) scanning (e.g. Støren et al., 2010), or spectral reflectance and colour imaging (Debret et al., 2010).
There are multiple sources of historical flood data (e.g. Brázdil et al., 2012), and depositories of historical flood data can be found in Brázdil et al. (2006a) and Kjeldsen et al. (2014). An overview of historical floods in Norway is available in Roald (2013). For quantitative analyses, it is nonetheless necessary to find evidence of historical flood stages, e.g. from flood stones or flood marks, and estimate flood discharge based on hydraulic calculations (Benito et al., 2015).
Systematic measurements of floods date back to Common Era (CE) 1870. Historical flood information in Norway is often available back to the 17th century; there is, however, scattered information on earlier floods, including one that occurred in the 1340s. This is different from paleoflood data in Norway, which typically cover the Holocene period (11 700 years) and extend all the way until the present day. The difference in time periods covered by diverse data sources on past flooding highlights the potential of using historical and paleo flood data to both reduce estimation uncertainty of design floods with long return periods and to assess non-stationarities in floods. The paleo and historical flood information can be used – in combination with systematic data – to estimate design floods (see e.g. Engeland et al., 2018; Kjeldsen et al., 2014; Stedinger and Cohn, 1986). To include the paleo and historical information in flood frequency analysis, we also need to know all floods exceeding a fixed threshold during a specified time interval. Several studies demonstrate that, given that the fixed threshold is high enough, it is adequate to know the number of floods exceeding this threshold in order to improve flood quantile estimates (Engeland et al., 2018; Martins and Stedinger, 2001; Payrastre et al., 2011; Stedinger and Cohn, 1986). A Bayesian approach to flood frequency analysis with historical and paleodata sources was introduced by Stedinger and Cohn (1986) and Gaál et al. (2010). This approach allows, in a flexible way, the introduction of multiple fixed thresholds and data sources and is therefore well suited for combining systematic, historical, and paleo data in a joint flood frequency analysis.
When we predict flood frequency for the future, the standard assumption is
stationarity or, put differently, it is assumed that the period with instrumental data is representative of the future. In many cases, when the
analysis is based on flood data from a streamflow gauging station covering a
limited period, it is a robust assumption (Serinaldi and Kilsby, 2015).
However, in the face of expected changes in climate, it is useful to take
into account the risk of floods in the future (Hanssen-Bauer et al., 2017; Lawrence, 2020; Paasche and Støren, 2014). For Norway, tailored guidelines for adaption to future flood risk are provided by the Norwegian Center for Climate Services (
Since the historical and paleo data cover much longer time periods than streamflow data, they can be an excellent source of non-stationarity in actual flood sizes and the underlying flood-generating processes. One approach is to link the frequency of floods to the underlying climatic drivers (e.g. mean temperature, precipitation, and large-scale circulation patterns) (e.g. Gilli et al., 2013; Kjeldsen et al., 2014; Støren et al., 2012; Støren and Paasche, 2014). A major challenge when using paleo and historical flood information is precisely to disentangle non-stationarity in climatic drivers from non-stationarities caused by changes in land use and/or the “archiving processes” of the data. Changes in land use can, for instance, be related to farming practices and timber logging. Changes in the archiving process might be caused by changes in the perception threshold that depend on societal development (Kjeldsen et al., 2014; Macdonald and Sangster, 2017). Also, changes in the river channel might limit the possibility of estimating the magnitude of paleo and historical floods (Brázdil et al., 2011).
The primary objective of this paper is to combine systematic, historical, and paleo information in a flood frequency analysis in order to better
understand and predict changes in flood frequency and magnitude for Norway's
largest river, Glomma. In particular we want to explore
past variability in floods as reconstructed from lake sediment cores;
potential non-stationarity in our new paleoflood record and its potential connection to regional climate change; the added value of combining systematic, historical, and paleo flood data when estimating flood quantiles; and potential non-stationarities in design floods.
The unique contribution of this study is thus to combine three different
information sources in an attempt to improve flood frequency estimations and
better understand the underlying mechanisms that cause significant changes
in flood variability over time.
The target site for this study is the city of Elverum lying next to the Glomma River. A gauging station with an upstream catchment area of 154 450 km
The location of the streamflow gauging station at Elverum used for flood frequency analysis and the site for paleodata collection close to Kongsvinger.
Seasonality of Glomma's monthly streamflow
The catchment has several hydropower reservoirs with a total regulation capacity presently around 10 % of the average annual runoff. The first reservoir was built in 1913, and since 1937 this and other reservoirs have resulted in decreased flood sizes (Pettersson, 2000). The monthly flows during winter have increased and most flood peaks have decreased since 1937 (Fig. 2). The catchment has undergone noteworthy land-use changes during the last 400 years. In the 17th to 19th centuries, the forest areas were reduced due to mining, timber export, and farming practices. Since the beginning of the 20th century, the forest-covered areas have increased slightly, whereas the timber volume has increased substantially, mainly due to farming and forestry practices, e.g. reduced grassing of domestic livestock and forestation (Grønlund et al., 1999).
To establish a flood record covering most of the Holocene (
Study site for the paleodata. Map: the sediment cores were extracted from Lake Flyginnsjøen. The green arrows indicate the flow direction under normal conditions, whereas the dark red arrow shows the flow direction whenever there is a flood that exceeds 1500 m
Schematic model of the lakes involved, flood water levels, thresholds, and flood pathways (after Hegge, 1968). The example shows the observed water level exceeding the threshold during the flood in 1967 (2533 m
Bathymetric map of Lake Flyginnsjøen and the coring sites which were chosen at the deepest part of the lake, close to the inlet. Note that the inlet during bifurcation events is only around 30 m away from the permanent inlet.
Annual maximum flood at Elverum (station number 2.604) for the period 1872–1936 was used for the flood frequency analysis. For this period, we assumed that the flood data were not significantly affected by river
regulations (Pettersson, 2000). The mean annual flood for the period 1937–2019 is 1362 m
Systematic and historical flood data at Elverum. The systematic data from 1872 to 1936 were used for flood frequency analysis. After 1937, the floods are dampened by river regulations. The flood in 1967 reached 2533 m
Historical flood information back to 1675 is available as water levels marked at a flood stone in Elverum, located close to Klokkerfossen (“fossen” meaning waterfall) at the Norwegian Forest Museum (Fig. 7 and Tables 1 and 2). Table 1 lists the water levels and discharges for floods exceeding the 1967 flood marked on the flood stone which was erected in 1968. The water levels were carved into the stone in 1969 based on recommendations from NVE (Hegge, 1969); the 1995 flood was added later. There is another flood stone nearby at Grindalen (also shown in Fig. 7). It was erected as early as in 1792 in order to remember the floods of 1773 and 1789, which were large indeed.
Water levels at Elverum gauging station and at the flood monument from Hegge (1969). The various streamflow peaks are constructed based on the rating curve at gauging station 2.119 and rating curve period 1881–1970. The large floods in 1966, 1967, and 1995 were not included in this study. The flood events in italics are from the period with systematic streamflow measurements.
Information about large historical floods at Elverum.
Map on the left shows the locations of the flood stones and the
gauging station at Elverum (left, created at
The flood stone at Grindalen is 2 km upstream of the flood stone at Klokkerfossen, with the streamflow gauging station at Elverum in the middle.
A waterfall at Klokkerfossen is the controlling profile for the water levels
at all three locations. Hegge (1969) developed relationships between water
levels at the Elverum gauging station and the flood stone at Klokkerfossen
shown here in Table 1. The water levels at the Elverum gauging station were transformed to discharges by using the local rating curve, which assumes that the river profile has been relatively stable since CE 1675. In this
study, we included all floods exceeding the observed 1967 flood peak at 2533 m
Table 2 summarizes the available historic information and important sources
for these floods. The floods in 1675, 1717, and 1749 are all described in
Finne-Grønn (1921) and Otnes (1982), whereas information for the flood mark in 1724 is not found in any written source. Detailed information on water levels for floods prior to 1773 was estimated in the absence of
historical data. The water levels in 1773, 1789, 1827, and 1846 are all engraved in the flood stone in Grinsdalen and employed here as a basis for
calculating the water level at the Elverum gauging station and also for the
flood stone at Klokkerfossen. Having said that, we still included all flood water levels listed in Hegge (1969). More information on the historical flood of the Glomma River and at Elverum is provided by Finne-Grønn (1921),
Otnes (1982), and Roald (2013). During the period 1675–1870, we see that eight floods exceeded the observed 1967 flood peak at 2533 m
The largest historical flood in this region was Stor-Ofsen, which took place on 22–23 July 1789 when peak discharge reached 3900 m
Prior to Stor-Ofsen, there was a substantial amount of snow in the mountains, deep soil frost, and rainfall that had saturated the soil. During the actual flood event, warm and humid air masses from the south-east were blocked by colder air masses in the north-west, resulting in high rainfall over the entire region. The rainfall intensity peaked on 22 July. The flood started on 21 July in small brooks and culminated the following day (Østmoe, 1985). The main rivers at the bottom of the valleys rose to unprecedented levels, and the flood was also accompanied by numerous landslides. The water levels of this flood are known from several markings cut into rocks, and many flood levels were later transferred to monuments erected at locations near the major rivers (Engeland et al., 2018; Finne-Grønn, 1921; Otnes, 1982; Roald, 2013).
Descriptions of bifurcation events and lists of estimated flow volumes in
Glomma at Kongsvinger are found in Aano (2017), Pettersson (2001), Hegge (1968), and Reusch (1903). From 1851 to 2013, 79 events in 77 different
years were recorded. In 1957 and 1987 there were bifurcation events both in
the spring and in the autumn; 4 of the 79 events occurred during the autumn.
For the interval between 1953 and 2013, the same period that is covered by FLS113, there were 22 bifurcation events. The transferred volume for the
period 1851–2013 is presented in Fig. 8. The five years with the largest
transferred volumes are 1916, 1934 1966, 1967, and 1995, with corresponding peak floods at Elverum yielding 2892, 2963, 2600, 2533, and 3238 m
Transferred volume (M m
Two sediment cores were retrieved from Flyginnsjøen in 2013 (see Sect. 2.2). Coring sites shown in Fig. 5 were selected at the deepest part of the lake based on a bathymetric survey of the lake using a Garmin Fishfinder echo sounder. A 516 cm long sediment core was retrieved using a 110 mm diameter piston corer (FLP213) (Nesje, 1992). Since the piston corer may disturb sediment layers in the upper 15–20 cm, an HTH gravity corer (FLS113) (Renberg and Hansson, 2008) was used to retrieve a 18 cm core of the youngest sediments. Samples of 1 cm
The area between Vingersjøen and Flyginnsjøen (Fig. 4) is rich in glaciofluvial deposits easily remobilized whenever floods occur. Bifurcation events in Glomma cause precisely such a fundamental change in the erosion regime in this area, causing river flooding in a normally dry area (see Sect. “Bifurcation events”). The following calculations and interpretations are thus based on the assumption that bifurcations events can be recorded as a marked increase in minerogenic input to Lake Flyginnsjøen, redeposited from the pre-existing glaciofluvial deposits in the catchment.
To quantify the frequency of such events, a local peak detection algorithm was applied to parameters sensitive to changes in minerogenic input. Flood
deposits were defined as peaks in the measured parameters where (i) the measured concentration is higher than the two surrounding values, (ii) the
difference between the peak and the lowest value within a specified time
window (
To produce a Holocene flood record based on the sediment cores from Flyginnsjøen, depth in the core was transformed to time using Bacon age–depth modelling software (Blaauw and Christeny, 2011) (see Sect. 4.1.1), and frequency of events in a 50-year moving window was quantified. In order to test to what extent the lake sediment records reproduce modern and historical observations, identified flood layers were compared with instrumental streamflow data.
A generalized extreme value (GEV) distribution was invoked to establish a flood frequency model for floods at Elverum. The GEV distribution is shown
to be a limiting distribution for block maxima (Embrechts et al., 1997;
Fisher and Tippett, 1928; Gnedenko, 1943):
We used non-informative priors for the location and scale parameters (i.e. the location parameter and the log-transformed scale parameter were uniform). A normal distribution with standard deviation 0.2 and expectation 0.0 was used as the prior for the shape parameter
The likelihood for the systematic data is (see Gaál et al., 2010; Stedinger and Cohn, 1986)
We also need to include available knowledge on floods exceeding
The posterior distribution of the parameters was estimated using a Markov chain Monte Carlo (MCMC) method implemented in R package nsRFA (Viglione, 2012). To estimate return levels, we used the posterior modal values of the parameters. It poses a challenge to set the perception threshold
The plotting positions provided by Hirsch and Stedinger (1987) that build on the Cunnane plotting position (Cunnane, 1978) were used to plot the
empirical distribution of the observations. The exceedance probability
We applied a simple approach to get an estimate of the non-stationary
200-year flood during the recent 1000-year one using the paleorecord. In a first step the parameters
From the sediment core we estimated a time series of the probability of
exceedance
The shortest core (FLS113) is 18 cm long and represents the period 1953–2013 (see Fig. 11). The longest core (FLP213) is 516 cm long and represents the period approximately 0–10 300 years before present (present
The results from the XRF scan (Ti
Results from measured parameters in FLP213.
The correlation matrix (Table 3) shows strong (and significant) correlations
between K
Correlation between measured parameters in FLP213 (in bold) and
FLS113 (in italic). LOI, BW, DBD, MS, and the XRF data (K, Ca, and Ti) were measured in FLP213, whereas CT greyscale, MS, and the XRF data (K, Ca, and Ti) were measured in FLP113. LOI (%) indicates content of organic matter in the core; BW is the 8-bit (0–255) black–white values extracted from a photo of the core surface where 0 is black. CT greyscale is a 16-bit number indicating relative densities of the core; DBD is given in unit g cm
This core shows dark organic gyttja with light grey minerogenic layers,
similarly to FLS213. The minerogenic layers yield high values of
K
Results from high-resolution analysis of core FLS113.
Correlation coefficients between CT greyscale values, MS, K
To establish an age–depth relationship for the cores, sediments were subjected to lead dating (
Fallout radionuclide concentrations and chronology for FLS113 from Flyginnsjøen.
Age–depth model for FLP213
We used the concentrations of Ti
Transferred volume of the 23 bifurcation events in the period
1950–2013 CE (in blue) and the 24 identified flood layers (red) identified using XRF scans of Ti
From FLS113 we have established a link between dense, minerogenic sediment
layers and bifurcation events. We therefore assumed that the analyses of
FLP213 could be used to produce a time series of flood events covering the
last 10 300 years. Here we used the local peak detection algorithm presented above to identify sediment layers with high concentrations of K
In
A key observation in the Holocene flood frequency reconstruction is the large non-stationarity played out across multiple timescales. We observe that there are two major flood-rich periods during the Holocene (Fig. 13a). The first runs from 3800 to 2000 cal yr BP, when it ends abruptly. The second period extends from around 700 cal yr BP up to the present day. Looking at flood frequency over the most recent 1000 years (Fig. 13b), we observe significant internal variability within the flood-rich period. The period with the highest flood rates occurs in the 18th century but also in the 15th century. The data from FLP213 inform us that the flood event in 1789 is truly an anomaly, as is evident from the sheer amount of sediments deposited during this event (no other flood comes close), and it also yields the highest measured values of e.g. density (DBD) as well as magnetic susceptibility (MS) throughout the core (Fig. 9). It is therefore reasonable to assume that the 1789 CE flood was an extraordinary event, making it the largest during the entire time span of the record, i.e. 10 300 years.
Overview of the three data sources used for flood frequency analysis.
The sensitivity of flood frequency analysis to three different
combinations of systematic and historical flood data. Annual maximum floods
for the period 1872–1836 were used as systematic flood data. Nine historical
floods exceeding 2533 m
The flood quantiles combining the systematic, historical, and paleo data have been analysed in different but complementary ways. Table 6 provides an overview of the flood information related to the data source and the time
intervals they represent. The first step was to estimate the flood quantiles
using only systematic data. In the second approach we added all the
historical flood data. The smallest historical flood of 2533 m
The next step was to include the paleoflood information in the flood frequency analysis. We did this in two ways: (i) we combined the systematic
data and the paleodata and (ii) we combined systematic, historical, and paleo data. For the paleodata we used 1800 m
The sensitivity of flood frequency analysis to three different
combinations of systematic, paleo, and historical flood data. Annual maximum
floods for the period 1872–1836 were used as systematic flood data.
Paleofloods representing 208 events exceeding 1800 m
To achieve a non-stationary estimate of the design flood, we used the flood occurrence rate presented in Fig. 13 to estimate the 200-year flood in a moving time window as explained in Sect. 3.4.2. We used 1900 m
Non-stationary estimate of the 200-year flood for the recent 6000 years. The red lines indicate the estimated 200-year flood and the 95 % confidence intervals estimated using systematic streamflow observations.
The historical data applied in this study are marked as water levels at the flood stone at Elverum, and the associated flood discharges are estimated by Hegge (1969). An assumption for these estimates is that the river profile is relatively stable over the historical period and in particular that the large flood in 1789 CE did not cause any substantial changes. This is a reasonable assumption because, although four large floods occurred between 1781 and 1969 CE, only one rating curve is used for the period. The gauging station was moved around 660 m in 1969.
During the last decade or so lakes across Europe have been studied in detail
and high-resolution paleoflood records have been produced from both the
lowlands and the highlands (cf. Wilhelm et al., 2018). Unlike many of these studies, we have worked with lakes that
The first assumption is that all flood events recorded in Lake Flyginnsjøen are directly related to Glomma. We cannot completely rule out the possibility that minor floods in the local catchment of Flyginnsjøen occurred simultaneously with floods originating from Glomma
or even just within the very small catchment surrounding the lake due to
local rainstorm events. Given the heavy vegetation cover in the catchment of
Flyginnsjøen, its small size, and the low angles of the slopes leading into the lake, we deem the possibility of a local sedimentary imprint to be very low. This is supported by both XRF and MS data. The consistency in
bifurcation events causing peaks in concentration in both Ti
A second assumption is that the river channel and landscape geometry
controlling the bifurcation events have not changed over the approximately recent 10 000 years to the extent that it alters this interplay between a
flooding Glomma and the investigated lake. The current river geometry was
shaped by a glacial lake outburst flood (GLOF) some 10 000–10 400 years
ago with a peak discharge of more than 10
The resolution of the XRF signal is on average sub-annual, but because of the uncertainty in the age–depth we calculated flood rates, i.e. average number of flood events, for a moving 50-year window. Unlike the findings of Evin et al. (2019), and although the floods are of varying magnitude, there appears to be no systematic relationship between flood sizes and sediment thickness or volume except for the Stor-Ofsen event. This is probably explained by the fact that the sediment transport for individual floods will in part be deposited in the two preceding lakes (Vingersjøen and Tarven) buffering Flyginnsjøen (Fig. 4) but may also indicate that event-specific features such as ground frost or snow cover may regulate sediment availability.
The paleoflood data presented here document that the flood frequency is non-stationary during the last 10 300 years, being manifested on multiple timescales (Fig. 13). Non-stationarity is typically identified as quasi-cyclic flood-rich and flood-poor periods (for European studies, see e.g. Brázdil et al., 2005; Glaser et al., 2010; Hall et al., 2014; Jacobeit et al., 2003; Kundzewicz, 2012; Mudelsee et al., 2004; Swierczynski et al., 2013), where the flood-rich period may last for 50–60 years (e.g. Glaser et al., 2010). Over the instrumental and historical eras, floods in the Glomma catchment have mainly occurred in late spring (late May, early June) due to the sudden melting of large snow reservoirs following a steep rise in temperatures that often overlaps with persistent rain (Roald, 2013). Under the current climate conditions, the largest floods in the Glomma catchment are caused by (i) high winter precipitation and preferentially cold winters resulting in a large snow storage, (ii) a cold spring followed by a sudden increase in air temperature producing high melt rates, and (iii) large amounts of widespread precipitation combined with snowmelt (Vormoor et al., 2016). Importantly, for these spring-snowmelt-triggered floods, the soils are either frozen and/or already saturated with moisture channeling shallow sub-surface flow and overland flow resulting in a fast discharge response to snowmelt and rain. Based on these observations, we hypothesize that on decadal to centennial timescales, increasing flood sizes can be explained by increasing precipitation, in particular during winter and spring, and cool winter temperatures. Increasing spring and summer temperatures might potentially lead to increasing flood sizes, but this effect depends strongly on the snow storage available for melt.
In Figs. 17 and 18 we compare the flood frequency reconstruction from Flyginnsjøen to several climate reconstructions representing temperature and precipitation on a centennial- to decadal-scale variability. In Fig. 17, the flood frequency is compared to regional summer temperature reconstructions (Moberg et al., 2005), whereas it is compared to local records of glacier variability (upper panel), a flood index (second panel), and local July temperature (third panel) in Fig. 18. No continuous reconstructions of winter precipitation are available for this region; however, the glacier growth in Scandinavia is primarily driven by summer temperatures and winter precipitation, and the reconstructed flood record is therefore compared to glacier variability in Rondane in the upper Glomma catchment. Low values of the flood index produced by Støren et al. (2012) reflect periods with relatively high flood frequency in eastern Norway. We observe co-variability between the reconstructed flood frequency in Flyginnsjøen and several of the climate reconstructions, which may indicate that the non-stationarity of flood frequency is, to a large degree, related to non-stationarities in climate. The data from Flyginnsjøen show, for instance, two distinct intervals with high flood frequency during the LIA, both played out on centennial timescales. Since 1850 there has been a steady increase in summer temperature followed by a reduction in flood frequency. Enhanced flooding during the LIA is observed in other lake studies from eastern Norway as well, including Atnasjø (Nesje et al., 2001), Butjønna (Bøe et al., 2006), Meringdalsvannet (Støren et al., 2010), and also the Grimsa River in the headwater of Glomma (Killingland, 2009).
Flood frequency in Glomma (blue bars) and 30-year moving average Northern Hemisphere summer temperature anomaly from Moberg et al. (2005).
Flood frequency in Glomma (red) with a 500-year running average, reconstructed summer temperature from Brurskardtjørni, southern Norway (Velle et al., 2010) (green) with a five-point running average, and flood index (blue) with a five-point running average (Støren et al., 2012) showing the relative distribution of the flood recurrence rate over southern Norway. Glacier activity at Skriufonn, Rondane, southern Norway (Kvisvik et al., 2015) (black and purple).
Another period with heightened flood activity occurred roughly between 4000 and 2000 years ago. The increase in flood frequency in Glomma during this period, and also during the LIA interval, coincides with a recorded decrease in summer temperature at Bruskardstjørni in eastern Jotunheimen (Velle et al., 2010) and increasing glacier growth in Rondane (Kvisvik et al., 2015), the mountainous source area of Glomma (Fig. 18). Multi-decadal periods are typically superimposed on centennial trends, as is the case for both these two flood-rich intervals. The near absence of floods prior to 4000 years ago is another recurring feature in all flood records from eastern Norway (e.g. Støren et al., 2016). Locally, it seems plausible that the effect of raising the 0-isotherm with 100–300 m altitude, the effect of a warmer summer season, will significantly change the potential storage of snow (Støren and Paasche, 2014).
The observed changes in flood frequency occurring during both the LIA and the first half of what sometimes is called the Neoglacial era (4000–2000 years ago) can thus, at least partially, be explained by the combined effect of the flood-generating processes (cf. Vormoor et al., 2016). The near absence of floods prior to the onset of the Neoglacial, when summer temperatures were ca. 1
In large catchments where snowmelt is the primary flood-generating process, it is suggested that we may see smaller flood sizes for eastern Norway according to Lawrence (2020). For small catchments, in western Norway, where rain-generated floods already dominate, floods are expected to increase. Cooler temperatures, especially in summer and spring, are likely to delay the melting of the snow cover – a scenario increasing the probability of a sudden warming simply because it occurs later in the season.
The increase in flood frequency commencing at ca. 4000 yr BP is a reoccurring feature not only in Europe, but also in parts of the USA (Paasche and Støren, 2014). This hints at a large-scale change in the climate system at the time, with implications for both atmospheric circulation patterns and temperature trends. This major climate shift recorded in Europe is noteworthy because the flood seasonality is different across such a large area for many reasons, including the varying altitudinal differences. In high-lying areas in Austria (north of the Alps, Swierczynski et al., 2013) and in the central Alps (Switzerland and northern Italy, Wirth et al., 2013), floods start to increase, as in eastern Norway, rapidly just after 4000 years ago and remain on average high until 2000 years ago. Studying the relative distribution of floods in Norway, Støren et al. (2012) suggest that the long-term trends in the floods are dependent on changes in the distribution of winter precipitation related to semi-permanent shifts in atmospheric circulation patterns and that an anomalously strong meridional component in the atmospheric circulation pattern is linked to floods in eastern Norway. Over the time period between the two flood-rich periods in Glomma (ca. 2000–1000 yr BP), Støren et al. (2012) recorded a westward shift in the flood frequency likely caused by reduced precipitation in the eastern areas (Fig. 18).
There are also potential catchment feedback mechanisms, not necessarily related to climate, that can both dampen and boost the flood patterns. Humans can potentially influence the landscape by forest clearing, which would alter sediment availability and runoff patterns as well as change the overall buffering capacity.
The flood-rich period occurring between 2500 and 4000 yr BP coincidences largely with the Bronze Age (2500–3700 BP), when settlements and farming expanded in Norway (Hjelle et al., 2018), but whether this early colonization impacted flood patterns remains an open question. Its worthwhile noticing that this interval with increased flooding is also recorded in other lake sediment records from southern Norway (Støren et al., 2010; Bøe et al., 2006) which were only marginally impacted by human activity, if at all (shown in Fig. 18), by farming. We therefore argue that the effect of land use cannot be the main explanation for the observed changes in flood frequency during this period. A similar conclusion was reached by Shubert et al. (2020), who show that logging and agricultural activities around the Mondsee, Austria, were low during flood-rich periods and that the flood record reflects climate variability rather than human activity in the catchment.
In more recent times, deforestation is a candidate that potentially could
help explain the increase in flood frequency after 1600 CE. The mining
industry that started in Norway in the late 16th century required a
large amount of timber, which resulted in widespread deforestation in Glomma's upper catchment. This removal of woodland cover may have influenced
the local erosion and sediment transport of the upstream Glomma catchment,
but because this area represents only a fraction of the total catchment area, we think that these “excess sediments” would be diluted downstream.
Another relevant point here is that the flat downstream gradient of the
Glomma River potentially causes sediment deposition long before it reaches the bifurcation point at Kongsvinger. A final point is that the sediment source for the flood layers deposited in Flyginnsjøen is suggested to be
mainly local, and the area around the lake and the location of the bifurcation events themselves were not subject to removal of woodland in this
period. It is possible that the removal of woodland amplified the size (and
frequency) of floods since forests, in most cases, reduce flood peaks. This, however, requires more regional and systematic vegetation change than that related to mining in the upper Glomma catchment to affect the 154 450 km
The non-stationarity in flood frequency is a major challenge when estimating flood quantiles used for land planning and design of infrastructure given that one needs to predict how the flood frequency will evolve over the lifetime of the construction; e.g. for bridges it is 100 years (Koh et al., 2014). Milly et al. (2008) argued that “stationarity is dead” and that it is necessary to account for non-stationarity in order to avoid underestimation of risks based on design floods.
Conversely, Serinaldi and Kilsby (2015) posited that “stationarity is undead” because a stationary model is robust and can be a useful reference/benchmark. Accounting for uncertainty in a stationary model can be as important as including non-stationarity within a risk assessment framework. A non-stationary model introduces more parameters and, thereby, in most cases increases the estimation uncertainty. An additional challenge when applying a non-stationary model for design flood estimation is to project the flood frequency into the future.
The paleoflood data presented here suggest that the flood frequency is non-stationary and that there are indeed flood-rich and flood-poor periods (Fig. 13). Since design flood estimates are used for assessing average risk over the lifetime of a construction, it is desired that design flood estimates are stable over time and not sensitive to quasi-cyclic variations in flood sizes on annual to decadal timescales. It is, however, important to account for trends or shifts in flood frequency. Macdonald et al. (2014) show that on centennial timescales, the effect of cyclic variations in short systematic records can effectively be removed by a temporal extension of flood time series using historical information. Data from Flyginnsjøen and historical data reveal that a quasi-stationary period can be identified at centennial timescales but not on a sub-millennium timescale where major shifts in flood frequency are identified (Fig. 13).
In this study, we firstly used the stationarity assumption and evaluated several possible ways to combine the three data sources within a stationary framework. The results in Figs. 14 and 15 show that the design flood estimates are sensitive to how we combine the systematic, historical, and paleo flood data. We used 65 years of systematic data covering the period 1872–1936 CE for which we assume that the effect of river regulation is negligible. Adding the historical data from the flood stone covering the period from 1653 to 1871 CE substantially increased the estimates of the flood quantiles and slightly reduced the estimation uncertainty (Fig. 14).
The paleoflood time series provided here suggests that the flood frequency during the historical period is non-stationary where the 18th century was an extremely flood-rich period (Fig. 13) and that the 1789 CE flood was an exceptional flood during the 10 300 years covered by the sediment core. Based on this paleo information, we used historical data from the 19th century and added the 1789 CE flood by assuming it was the largest flood over a period of 10 300 years. This slightly reduced the flood quantile estimates as compared to using all historical information and substantially reduced the estimation uncertainty (Fig. 14). These results show that for the site at Elverum, we should be careful when including historical flood information from the flood-rich period in the 18th century.
As a next step, we added the paleoflood data representing 572 years (i.e. 1300–1871 CE). This resulted in negligible differences in flood quantile and uncertainty estimates (Fig. 15) indicating that the information content in the paleodata alone can be small. A possible explanation is the combination of the relatively low threshold (according to Fig. 15 it is around a 5-year flood) and that we only had information on the number of flood events. Both Macdonald et al. (2014) and Engeland et al. (2018) show that the information content is low when the threshold for historical floods is too low.
In a final step we used the flood rate from the sediment core as a key to explore non-stationarity of the design flood estimates, exemplified by the 200-year flood (Fig. 16). We could see important variation during the recent 6000 years. The 200-year flood was estimated to be around 23 % higher during the flood-rich periods in the 18th century and 20 % lower during the warmest period. The high values for the 200-year flood during the 18th century is confirmed by the historical data. This variation in design floods is, interestingly within the range seen in recent studies on climate change impacts on floods in Norway (Lawrence, 2020). For a future climate that is expected to be warmer, the design flood might be expected to decrease. Furthermore, this shows that the most interesting information we could get from the sediment core was the non-stationarity in floods.
In this study we have (i) compiled historical flood data from the existing literature, (ii) presented an analysis of the sediment core extracted from Lake Flyginnsjøen in Norway including results of XRF and CT scans plus
MS measurements and used these data to estimate flood frequency over a
period of 10 300 years, and (iii) combined flood data from systematic
streamflow measurements, historical sources, and lacustrine sediment cores for estimating design floods and assessing non-stationarities in flood
frequency at Elverum in the Glomma catchment located in eastern Norway. Our
results show the following.
Based on detailed analysis of lake sediments that trap sediments whenever the Glomma River exceeds a local threshold, we could estimate flood frequency in a moving window of 50 years throughout the last 10 300 years. The paleodata show that the flood frequency is non-stationary across timescales. Flood-rich periods has been identified, and these periods corresponds well to similar data in eastern Norway and also in the Alps such as the increase around 4000 years ago. The flood frequency can show significant non-stationarities within a flood-rich period. The most recent period with a high flood frequency was the 18th century, and the 1789 flood (Stor-Ofsen) is probably the largest flood during the entire Holocene. The estimation of flood quantiles benefits from the use of historical and paleo data. The paleodata were in particular useful for evaluating the historical data. We identified that the 1789 flood was the largest one for the recent 10 300 years and that the 18th century was a flood-rich period as compared to the 20th and 19th centuries. Using the frequency of floods obtained from the paleoflood record resulted in minor changes in design flood estimates. We could use the paleodata to explore non-stationarity in design flood estimates. During the coldest period in the 18th century, the design flood was up to 23 % higher than today, and down to 30 % lower in a warmer climate ca. 4000–6000 years ago.
This study has demonstrated the usefulness of paleoflood data, and we suggest that paleodata have a high potential for detecting links between climate dynamics and flood frequency. The data presented in this study could be used alone or in combination with paleoflood data from other locations in Norway and Europe to analyse the links between changes in climate and its variability and flood frequency.
Systematic flood data are available from the national hydrological database at the Norwegian Water Resources and Energy Directorate. The data from the scanning of the sediment cores are available upon request to the authors.
The study was designed and planned by AA, KE, IS, and ES. IS and ES carried out the lake coring and the field work. AA, IS, ØP, and ES all contributed the scanning and analysis of the sediment cores. AA, KE, and ES contributed to systematization of historical and systematic flood data and the flood frequency analysis. AA prepared Figs. 1 and 3. ES and ØP prepared Fig. 4. ES prepared Figs. 5, 9–11, and 18. KE prepared Figs. 2, 6–8, and 12–17. KE prepared the manuscript with contributions from all the co-authors.
The authors declare that they have no conflict of interest.
We would like to extend our thanks to Svein Olaf Dahl, Nils Roar Sæltun, and Chong-Yu Xu, who co-supervised the master projects of Ida Steffensen (Dept. of Geography, UiB) and Anna Aano (Dept. of Geoscience, UiO) that create the basis for this study, and Karoline Follestand and Martin Tvedt, who assisted in coring Lake Flyginnsjøen. This study became part of the Hordaflom project which is funded by RFF-Vest, Hordaland County. Øyvind Paasche is grateful to project ACER, funded by the Research Council of Norway. All core samples, apart from the dating, were measured at the Earth Surface Sediment Laboratory (EARTHLAB) at the Department of Earth Science, University of Bergen.
This research has been supported by the Research Council of Norway (grant nos. 812957 and 226171) and the RFF Vestlandet (grant no. 269682).
This paper was edited by Roger Moussa and reviewed by Daniel Schillereff and one anonymous referee.