Hydrological Sciences Journal

ISSN: 0262-6667 (Print) 2150-3435 (Online) Journal homepage: http://www.tandfonline.com/loi/thsj20

Trend detection in river flow series: 2. Flood and low-flow index series / Détection de tendance dans des séries de débit fluvial: 2. Séries d'indices de crue et d'étiage Cecilia Svensson , W. Zbigniew Kundzewicz & Thomas Maurer To cite this article: Cecilia Svensson , W. Zbigniew Kundzewicz & Thomas Maurer (2005) Trend detection in river flow series: 2. Flood and low-flow index series / Détection de tendance dans des séries de débit fluvial: 2. Séries d'indices de crue et d'étiage, Hydrological Sciences Journal, 50:5, -824, DOI: 10.1623/hysj.2005.50.5.811 To link to this article: http://dx.doi.org/10.1623/hysj.2005.50.5.811

Published online: 15 Dec 2009.

Submit your article to this journal

Article views: 1040

View related articles

Citing articles: 66 View citing articles

Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=thsj20 Download by: [37.44.207.3]

Date: 14 January 2017, At: 23:36

Hydrological Sciences–Journal–des Sciences Hydrologiques, 50(5) October 2005

811

Trend detection in river flow series: 2. Flood and low-flow index series

CECILIA SVENSSON1, ZBIGNIEW W. KUNDZEWICZ2,3 & THOMAS MAURER4 1 Centre for Ecology and Hydrology, Crowmarsh Gifford, Wallingford, Oxfordshire OX10 8BB, UK [email protected] 2 Research Centre of Agricultural and Forest Environment, Polish Academy of Sciences, Bukowska 19, 60-809 Poznań, Poland 3 Potsdam Institute for Climate Impact Research, Telegrafenberg, D-14412 Potsdam, Germany 4 Global Runoff Data Centre, Federal Institute of Hydrology, Am Mainzer Tor 1, D-56068 Koblenz, Germany

Abstract Major floods in Europe and North America during the past decade have provoked the question of whether or not they are an effect of a changing climate. This study investigates changes in observational data, using up to 100-year-long daily mean river flow records at 21 stations worldwide. Trends in seven flood and low-flow index series are assessed using Mann-Kendall and linear regression methods. Emphasis was on the comparison of trends in these flow index series, particularly in peak-overthreshold (POT) series as opposed to annual maximum (AM) river flow series. There is a larger number of significant trends in the AM than in the POT flood magnitude series, probably relating to the way the series are constructed. Low flood peaks occurring at the beginning or end of a time series with trend may be too low to be selected for the POT analysis. However, one peak per year will always be selected for the AM series, making the slope steeper and/or the series longer, resulting in a more significant trend. There is no general pattern of increasing or decreasing numbers or magnitudes of floods, but there are significant increases in half of the low-flow series. Key words annual maximum series; climate change; flood; linear regression; low flow; Mann-Kendall test; peak-over-threshold series; trend analysis

Détection de tendance dans des séries de débit fluvial: 2. Séries d’indices de crue et d’étiage Résumé Les crues majeures de cette dernière décennie en Europe et en Amérique du nord ont posé la question de savoir si elles sont ou non un effet d’un changement climatique. Cette étude examine les changements dans des données d’observations, en s’appuyant sur 21 stations réparties dans le Monde, qui disposent de chroniques de débits fluviaux moyens journaliers dont la longueur va jusqu’à 100 ans. Les tendances sont analysées, par les méthodes de Mann-Kendall et de régression linéaire, dans des séries de sept indices de crue et d’étiage. La priorité a porté sur la comparaison des tendances mises en évidence pour ces séries d’indices, en particulier pour les séries de pics seuillés (PS) par rapport aux séries de débits fluviaux maximum annuels (MA). Les tendances significatives sont plus nombreuses pour les séries de MA que pour les séries de PS, ce qui est probablement lié à la façon dont les séries sont construites. Les petits pics de crue qui apparaissent au début ou à la fin d’une série temporelle avec tendance sont peut-être trop petits pour être sélectionnés lors de l’analyse de PS. Par ailleurs, un pic sera toujours sélectionné pour chaque année dans les séries de MA, rendant la pente plus raide et/ou la série plus longue, et produisant donc une tendance plus significative. Il n’y a pas de schéma général d’augmentation ou de diminution des nombres ou des amplitudes de crues, mais une augmentation significative apparaît pour la moitié des séries d’étiage. Mot clefs série de maximum annuels; changement climatique; crue; régression linéaire; étiage; test de Mann-Kendall; série de pics seuillés; analyse de tendance

Open for discussion until 1 April 2006

Copyright  2005 IAHS Press

812

Cecilia Svensson et al.

INTRODUCTION Major floods in Europe and North America during the past decade (e.g. Kunkel et al., 1994; CEH Wallingford/Met Office, 2001; Marsh & Bradford, 2003; Saurí et al., 2003; Kundzewicz, 2004) have provoked the question of whether or not they are an effect of a changing climate. Results from hydrological models that use output from general circulation models often suggest that river flows will increase in a greenhouse gas-induced warmer future climate (e.g. Miller & Russell, 1992; Nijssen et al., 2001; Reynard et al., 2001; Milly et al., 2002). Middelkoop et al. (2001) found that the flood risk in the Rhine basin can be expected to rise in winter, whereas at the same time summer droughts may become more severe. The Intergovernmental Panel on Climate Change’s Third Assessment Report (IPCC, 2001) concludes that an increasing body of observations gives a collective picture of a warming world and other changes in the climate system. Observational evidence suggests that it is likely that heavy precipitation events have increased at midand high northern latitudes, whereas the frequency and severity of droughts in some regions of Asia and Africa have also increased. The mechanism whereby an increase in greenhouse gases in the atmosphere would produce global warming is through an increase in downwelling infrared radiation (Trenberth, 1998). This would not only increase surface temperatures, but also enhance the hydrological cycle, as much of the heating at the surface goes into evaporating surface moisture. With higher temperatures in the atmosphere, the water-holding capacity also increases, and, together with an increase in evapotranspiration, this suggests that the actual atmospheric moisture content would increase. Trenberth (1998) argues further that, in general, the atmospheric moisture increase is likely to result in heavier rainfall and therefore also in an increased flood risk, although locally the impacts of climate change may be different. At the same time, the increase in evapotranspiration—and in some areas decrease in precipitation—may lead to longer and more severe droughts in the dry season. Trend analysis can be used to investigate whether there is any support for increases in river floods and droughts in observational flow data. By using observations rather than model output, uncertainties inherent in the modelling procedure, such as simplifying assumptions and concepts, are avoided. However, using real data involves other problems, relating chiefly to data quality but also to quantity. Trend analysis requires long records, preferably in excess of about 50 years (Kundzewicz & Robson, 2000), to distinguish climate change-induced trends from climate variability. There are not sufficient metadata available for the flow series used in the study, which means that any trends may include the effects of changes in land use and flow attenuation due to reservoirs, as well as any impacts of climate change. Flood trend studies tend to focus on trends in the annual maximum flood series. With the advent of computers and digitized observations, it has become easier to extract more information for use in hydrological analysis (e.g. flood frequency analysis and trend analysis) than is provided solely by one value per year. For example, a peakover-threshold approach (e.g. Stedinger el al., 1993; Robson & Reed, 1999) selects all floods above a certain threshold that occur in an entire flow record, thus also allowing an estimate of the trend in the frequency (counts) of floods rather than just their magnitude. Copyright  2005 IAHS Press

Trend detection in river flow series: 2. Flood and low-flow index series

813

Kundzewicz et al. (2004, 2005) investigated trends in annual maximum daily mean river flows at 195 stations with a global spread. Using a subset of 21 stations from these 195, the present study extends the analysis to explore whether the results are similar when using peak-over-threshold (POT) methods as opposed to using annual maximum river flows. Trends in both POT magnitude, as well as in number of POTs per year, are estimated. To give an indication of whether more/less flooding tends to be accompanied by more/less drought, trends in low-flow indices are also investigated. Kundzewicz et al. (2004, 2005) found some spatial continuity in the trends of annual maximum river flows. However, generally the results were too inhomogeneous, and the density of stations too low, for the subset of individual stations used in the current study to be regarded as representative for any particular region. A more detailed account of the study can be found in Svensson et al. (2004). TREND ANALYSES OF OBSERVED RIVER FLOODS AND LOW FLOWS Because of land-use changes, reservoir construction, and other local effects, there are seldom perfectly homogeneous spatial patterns emerging from regional studies of trends in floods. However, with some generalization, results from studies using daily mean flow data can be summarized as follows (significance at the 95% level): there is some evidence of decreasing trends in floods in western Canada (Adamowski & Bocci, 2001; Burn & Hag Elnur, 2002; Kundzewicz et al., 2004), whereas most of the USA have few significant trends, and the ones observed are of varying direction (Douglas et al., 2000; Lins & Slack, 1999; Kundzewicz et al., 2004). In Europe there is some evidence of significant positive flood trends in northern Scandinavia (Lindström & Bergström, 2003, 2004; Kundzewicz et al., 2004), but no regional flood trends could be found in the UK (Robson et al., 1998). Although there are significant trends at a quarter of the stations in Central Europe (Kundzewicz et al., 2004), they are both positive and negative. Because droughts tend to be longer lasting than floods, data of daily to monthly resolution are used for low-flow studies. Douglas et al. (2000) and Lins & Slack (1999) found increasing trends in low flows from the midwest towards the northeast of the USA. Lins & Slack (1999) also found that the annual median streamflow is increasing, whereas floods are neither increasing nor decreasing, leading them to conclude that the conterminous USA appears to be getting wetter, but less extreme. Adamowski & Bocci (2001) found mainly increasing trends in the western regions of Canada. Using 600 daily streamflow records in Europe, Hisdal et al. (2001) conclude that it was not possible to establish that drought conditions in general have become more severe or frequent. Cluis & Laberge (2001) investigated trends in the AsiaPacific region (including Oceania and the vast majority of Asia). Most areas do not exhibit consistent trends. However, in Central and Far-East Asia, rivers to the north (between the 50th and 75th parallels) exhibit upward trends, whereas more southern stations (around the 45th parallel) show downward trends. DATA Daily mean river flow data at 21 stations with a global distribution were used for the study (Table 1, Fig. 1). Data were obtained from the Global Runoff Data Centre Copyright  2005 IAHS Press

814

Cecilia Svensson et al.

Table 1 General information about the 21 daily mean river flow gauges used in the study. In the column for amount of missing data, 0 means the record is complete, and 0.00 means that there is less than 0.005% data missing. Country* River and station location Longi- LatiArea First and Record Amount tude tude (km2) last years length missing (°E) (°N) (years) (%) ML Niger at Koulikoro –7.55 12.87 120 000 1907–1987 81 0.10 ZA Mtamvuna at Gundrift 29.83 –30.73 715 1956–2000 45 1.36 RU Selenga at Mostovoy 107.48 52.03 440 200 1936–1999 64 0 RU Ob at Salekhard 66.53 66.57 2 949 998 1954–1999 46 0 TH Chao Praya at Wat Pho 100.19 15.17 120 693 1950–1999 50 0.83 Ngam (Ban Re Rai) 3206720 VE Orinoco at Puente –63.6 8.15 836 000 1926–1989 64 0 Angostura 3512400 GF Maroni at Langa Tabiki –54.43 4.98 60 930 1952–1995 44 0 4113300 US Red River of the North –97.03 47.93 77 959 1904–1999 96 0 Grand Forks, N.D. 4116300 US Clearwater River at –116.82 46.44 24 786 1926–1999 74 0 Spalding, ID 4148051 US James at Cartersville, VA –78.09 37.67 16 205 1900–1999 100 0 4150503 US Brazos River at Seymour, –99.27 33.58 40 243 1924–1999 76 0 TX 5202065 AU Styx River at Jeogla 152.16 –30.59 163 1919–1992 74 0.24 5204105 AU Murrumbidgee River at 149.09 –36.17 1891 1927–2000 74 0 Mittagang Crossing 5302250 AU Thomson River at Cooper 146.43 –37.99 906 1956–2001 46 0 Creek 5608024 AU Fitzroy River at Fitzroy 125.58 –18.21 45 300 1956–1999 44 0 Crossing 6142100 CZ Morava at Moravicany 16.98 49.76 1559 1912–2000 89 0.19 6335125 DE Kinzig at Schweibach 8.03 48.39 954 1921–2000 80 0 6545200 SI Krka at Podbocje 15.46 45.86 2238 1933–1999 67 0.00 6609400 GB Avon at Evesham –1.94 52.09 2210 1937–1999 63 0 6731300 NO Etna at Etna 9.43 60.93 557 1920–2000 81 0 6855100 FI Vantaanjoki at Oulunkyla 24.98 60.23 1680 1937–2001 65 0 (near the mouth) * Abbreviations of country names: AU Australia, CZ Czech Republic, DE Germany, FI Finland, GB Great Britain, GF French Guiana, ML Mali, NO Norway, RU Russia, SI Slovenia, TH Thailand, US United States, VE Venezuela, ZA South Africa. GRDC station number 1134100 1160650 2907400 2912600 2964130

(GRDC, 2003) in Koblenz, Germany. The selected records are a subset of the 195 records used by Kundzewicz et al. (2005). The 21 stations were selected to obtain an even geographic coverage worldwide of long records with few missing data. The record lengths vary between 44 and 100 years, with an average of 68 years. Unfortunately, the GRDC does not receive information about any changes to the gauging stations or the physical conditions in the catchments, so the suitability of stations for this kind of analysis could not be assessed. Smaller catchments are less likely to be affected by anthropogenic activities. However, especially in Africa, Asia and South America there were not many long time series with few missing data available, so some large catchments have been included in the study.

Copyright  2005 IAHS Press

Trend detection in river flow series: 2. Flood and low-flow index series

815

Fig. 1 Location of the 21 daily mean river flow stations.

METHODS The analysis was carried out using the HYDROSPECT software (Radzeijewski [sic] & Kundzewicz, 2004b). Seven different index series of river flow, such as for example the annual maximum river flow (resulting in a series consisting of one value per year), were extracted from the daily mean river flow series for the 21 stations. Two tests for trend, linear regression and the Mann-Kendall test, were then applied to the index series. The low-flow series show serial correlation over several years, which means that theoretical, parametric, formulae for estimation of trend significance as used in the accompanying paper (Kundzewicz et al., 2005), are not suitable for this study. Instead, significance levels are estimated using a block bootstrapping method. Flow indices Five different indices were used to describe the characteristics of the upper end of the flow regime, i.e. the floods. The first of these is the annual maximum daily mean river flow (Ann. max.), which was used by Kundzewicz et al. (2004, 2005) for the analysis of trends in river floods at 195 stations worldwide. In flood-rich years, the annual maximum series will include only the largest of several large observed flows, whereas in flood-poor years a small river flow will be selected that may not necessarily be characterized as a flood at all. One way of representing high river flows in a record, regardless of when they occur, is to use a peak-over-threshold approach (e.g. Stedinger et al., 1993; Robson & Reed, 1999). Peak-over-threshold (POT) series consist of a series of independent daily mean river flows that exceed a certain threshold (this threshold is the same throughout the time series). The POTs have to be proper peaks, i.e. the river flow both before and after the peak has to be lower than at the peak itself. Two POT indices describing flood magnitude were used; the POT1 magnitude (POT1 mag.) and the POT3 magnitude (POT3 mag.) series. The magnitude of the threshold Copyright  2005 IAHS Press

816

Cecilia Svensson et al.

was set so that on average one and three POT events, respectively, were selected per year. The peaks in a POT series were considered to be independent of each other if they were separated by a particular time interval. If two peaks occurred too close to each other, the smaller peak was dropped from the POT index series. After inspection of the time series, the time interval between peaks was set to at least five days for catchments with areas 100 000 km2. These separation times generally allow for the flow to recede appreciably between peaks. However, individual flood peaks are less pronounced on large catchments with a strong seasonal component, notably at station 3206720 (Orinoco at Puente Angostura), but also at 1134100 (Niger at Koulikoro). For these two catchments three peaks per year on average could not be extracted for an independence criterion of 20 days. Because no other suitable stations were available in these regions, it was decided to keep these stations in the study, and bear in mind the difference in number of peaks when interpreting the results. On average 2.1 and 1.5 peaks per year, respectively, for stations 1134100 and 3206720, were used for the POT3 series. The frequency (annual counts) of flood events can be described by counting the number of POTs occurring in each year. Two such flood frequency indices were used: the POT1 frequency (POT1 freq.) and the POT3 frequency (POT3 freq.). These annual frequency series were derived from the corresponding POT magnitude series. The two POT1 series describe the magnitude and frequency of the most extreme floods, whereas the two POT3 series characterize the behaviour also of the more moderately sized floods. Two low-flow indices were used to describe the lower end of the flow spectrum: the series of annual minimum 7-day (Min. 7-day) and 30-day (Min. 30-day) mean river flow. Particularly the 7-day duration is commonly used for low-flow analysis (e.g. Gustard et al., 1992). Trend detection Two different methods were used to estimate whether there is a significant positive or negative trend in the river flow index series. Linear regression fits a regression line to the series, and the slope describes whether the trend is strong or not. The null hypothesis is that the slope of the line is zero. Because the linear regression is applied directly to the index series, rather than to ranks, it is very good for visual presentation. However, the linear regression method works best when the residuals have a normal distribution, and it is very sensitive to outliers in the data. By ranking the observations and applying the distribution-free and non-parametric Mann-Kendall test, a more robust measure of trend is obtained (although, if a distribution assumption holds, a distribution-specific test is generally more powerful). For both the linear regression and Mann-Kendall tests significance levels were estimated using a distribution-free resampling method (see next section). Both trend tests test for inhomogeneities in a location parameter, e.g. the mean for the linear regression. However, they will also capture inhomogeneities resulting from changes in both the location and the variability. Kundzewicz & Robson (2000, 2004) Copyright  2005 IAHS Press

Trend detection in river flow series: 2. Flood and low-flow index series

817

and Radziejewski & Kundzewicz (2004a) describe and discuss methods for trend detection in more detail. Block bootstrapping for estimation of significance levels The 10% significance level (for a two-sided test) used by Kundzewicz et al. (2005) for the accompanying study on trends in annual maximum flood series, was adopted for presentation purposes also in this paper. Because the statistical distributions of the index series are not necessarily known, and the observations may not be independent and identically distributed, significance levels were estimated using a block bootstrapping method (e.g. Efron, 1979) rather than through theoretical formulae. The independence criterion is frequently broken as there is serial dependence from one year to another, mainly in the low-flow index series but also for some of the flood index series. The POT magnitude series may exhibit seasonality, i.e. the peaks may not be identically distributed. Bootstrapping is based on the generation of many new data sets, resamples. The original sample of observations is used as the distribution from which the resamples are chosen randomly with replacement, i.e. with each observation being returned to the original sample after it has been chosen, so that it may be chosen again. A large number of data sets are generated and a test statistic, in this case the estimate of the trend, is calculated for each of these new data sets. This provides a sample of the trend estimate that would occur for a range of situations, as the trend is calculated from some resamples that have a strong positive trend (i.e. many low observations at the beginning, and high observations at the end, of the resampled series), from some resamples without much trend, and from some resamples with a strong negative trend. Seasonality can be taken into account by using annual, or multiples of annual, blocks of data for the bootstrapping. Autocorrelation analysis was carried out for all the annual series to establish how long the blocks would need to be, to take into account the serial correlation of the series. It revealed that five-year blocks would be sufficient for most of the series. For the few cases where autocorrelation was significant for more than a 4-year lag, the significance levels for the trend analysis were estimated using longer blocks. When a longer block size was needed for the annual maximum flow series, the significance levels of trends for the POT magnitude series were also calculated using the larger blocks. After some initial tests, it was deemed suitable to make 2000 resamples per station and index to obtain good stability in the significance level estimates. The 2000 trend estimates of the resamples were subsequently ranked in descending order. If the trend of the original data series was outside the 5% and 95% points of the ranked trends of the resamples (i.e. the 100th and 1901st ranked values), the trend of the original data series was considered to be significant at the 10% level. Trend index A trend index was derived for presentation purposes and to make possible a correlation analysis of the trends. The trend index, TI, ranges from –100 to +100, with negative Copyright  2005 IAHS Press

818

Cecilia Svensson et al.

values indicating a negative trend and positive values a positive trend. The higher the absolute value of TI, the higher the significance of the trend. The TI (in %) relates to the significance level, α (also in %), of the two-sided test as: for positive trends ì 100 − α TI = í î− (100 − α ) for negative trends Thus, for example, a negative trend significant at the 10% level will have TI = –90%. RESULTS AND DISCUSSION For each station, Table 2 shows the trend index (TI) of the trends in the seven flow index series estimated using linear regression and the Mann-Kendall test. When the trend is negative, the TI is shown in italic font. Absolute values of TI exceeding 90% are shown in bold font. These values correspond to trends significant at the 10% level. The similarity between the columns in Table 2 can be assessed using correlation analysis. This shows that there is generally good agreement between the outcome for the linear regression and Mann-Kendall methods, except when there are one or more outliers in the flow index series. This may result in the two methods showing trends of the opposite sign. Examples of this are the flood magnitude indices Ann. max. and POT3 mag. for station 1160650 (Table 2 shows the different trend signs; Fig. 2(a) displays the outliers). Because the results of the Mann-Kendall tests are more robust, they have been used for the presentation in the following discussion unless otherwise stated. There is also very good agreement between the columns for the two low-flow indices, and generally between the different flood indices. However, the column for POT3 freq. is not significantly correlated with either the POT3 or POT1 magnitude columns at the 10% level. This means that an increase (decrease) in the magnitude of floods is not necessarily associated with an increase (decrease) in the frequency with which they occur. On the face of it, this contradicts the findings by Robson & Reed (1999), who investigated trends in British river flow series, using annual maximum and POT magnitude and frequency series with on average one and three POTs per year. They noted that trends in the annual maximum series were often associated with trends in the flood counts series, both trends generally being positive. However, an inspection of the individual results of Robson & Reed (1999) reveals that significant magnitude and frequency trends sometimes had opposite signs. The two stations for which three peaks per year on average could not be extracted (1134100 and 3206720, with on average 2.1 and 1.5 peaks per year, respectively) do not show widely different behaviour for the POT3 series compared with other stations. For the two stations, a couple of non-significant POT3 magnitude and frequency trends are of opposite direction to the corresponding POT1 trends, but four to five other stations show the same behaviour. For both the flood magnitude and frequency index series, there are generally slightly more stations showing a significant negative trend than a significant positive trend. The Ann. max. index has the highest number of both negative and positive significant trends: 7 and 4, respectively (Fig. 3). A more significant trend may occur in Copyright  2005 IAHS Press

–54.79 –17.76 72.80 –98.17 52.79 32.85 99.77 –97.62 96.82 –99.77 57.99 –92.37 –91.20 86.38 78.73 97.37 –98.32 95.77 –89.83 –19.76

Mtamvuna at Gundrift

Selenga at Mostovoy

Ob at Salekhard

Chao Praya at Wat Pho Ngam (Ban Re Rai)

Orinoco at Puente Angostura

Maroni at Langa Tabiki

Red River of the North Grand Forks, N.D.

Clearwater River at Spalding, ID

James at Cartersville, VA

Brazos River at Seymour, TX

Styx River at Jeogla

Murrumbidgee River at Mittagang Crossing

Thomson River at Cooper Creek

Fitzroy River at Fitzroy Crossing

Morava at Moravicany

Kinzig at Schweibach

Krka at Podbocje

Avon at Evesham

Etna at Etna

Vantaanjoki at Oulunkyla (near the mouth)

1160650

2907400

2912600

2964130

3206720

3512400

4113300

4116300

4148051

4150503

5202065

5204105

5302250

5608024

6142100

6335125

6545200

6609400

6731300

6855100

Bootstrapping in 6-year blocks. 2 Bootstrapping in 10-year blocks. 3 Bootstrapping in 14-year blocks.

1

–78.831

Niger at Koulikoro

1134100

–24.91

–91.22

95.62

–97.72

97.20

45.65

84.80

–98.27

–98.22

63.29

–99.67

97.22

–95.92

97.92

13.46

60.59

–99.70

65.44

–30.10

3.19

–66.691

–75.53

–96.32

75.18

–82.13

94.92

77.53

72.80

66.69

–36.65

–1.32

–98.77

89.23

–93.47

98.22

39.65

15.96

–91.32

44.60

42.80

–83.63

–35.301

–86.28

–98.22

41.40

–64.89

98.57

52.84

52.74

73.98

0.92

45.55

–97.32

90.62

–88.73

98.70

48.45

32.65

–95.72

28.86

68.79

–70.16

–30.401

Flood magnitude: Ann. max. POT1 mag. LR M-K LR M-K

River and station location

GRDC station no.

–88.80

–83.88

93.27

–99.47

96.77

84.98

37.15

61.40

41.85

81.30

–98.97

93.47

–98.97

98.12

–12.10

–72.38

–92.42

20.31

18.76

–67.94

–81.181

–69.24

–81.73

70.38

–98.72

78.33

18.81

–48.20

–3.22

63.34

92.37

–87.88

82.53

–96.37

79.88

–17.86

–13.76

–65.64

33.85

–20.71

19.76

–74.431

POT3 mag. LR M-K

0.70

–78.83

79.93

–99.97

83.33

43.40

80.40

–98.22

–93.22

89.73

–99.82

79.63

–94.20

99.67

–31.80

56.87

–81.88

47.40

–30.60

57.49

–53.54

67.94

–68.49

88.93

–99.82

91.17

54.24

80.48

–97.97

–94.20

89.48

–99.77

88.13

–93.12

98.47

–6.42

56.52

–89.98

45.35

–47.15

71.30

–51.14

89.18

72.91

–86.13

30.10

–99.42

28.16

–69.24

93.72

–99.57

97.42

79.93

96.373

74.30

98.47

92.22

71.23

86.13

–89.73

44.60

99.70

–44.95

99.622

95.77

26.71

87.18

11.21

98.97

98.22

–78.68

69.19

99.52

87.68

97.323

72.18

98.97

90.22

97.57

75.93

–84.73

87.38

99.92

–46.60

99.772

91.32

11.56

88.73

10.86

99.52

96.52

–84.78

37.70

Low flow: Min. 7-day LR M-K

98.87

89.28

99.92

52.74

91.17

69.74

71.63

88.93

–95.12

–42.45

98.97

–78.28

98.722

98.20

22.60

80.53

–6.17

99.27

98.22

–67.94

64.79

99.77

93.77

99.87

58.84

96.42

73.80

98.12

81.13

–91.77

75.53

99.82

–72.58

99.122

93.92

12.86

79.63

–11.36

98.77

97.27

–75.68

46.70

Min. 30-day LR M-K

Copyright  2005 IAHS Press

89.93

–89.53

49.20

–98.62

45.12

–60.89

95.17

–99.82

–98.87

1

1

–98.72

–40.90

–99.92

21.71

–81.48

98.70

–38.50

56.24

25.91

–43.85

–92.87

24.81

–35.70

–99.97

22.51

–83.63

99.70

–46.35

63.21

25.76

–48.12

–93.77

40.75

92.70

Flood frequency: POT1 freq. POT3 freq. LR M-K LR M-K

Table 2 Trend indices (TI) associated with trends for seven river flow index series. Trends are estimated using linear regression (LR) and the Mann-Kendall test (M-K). Negative trends are shown as having a negative TI, and are in italic font. Absolute values of TI exceeding 90% are shown in bold font. These values correspond to trends significant at the 10% level.

820

Cecilia Svensson et al.

Fig. 2 Observations and linear regression trends in the annual maximum daily mean flow series (Ann. max.), peak-over-threshold magnitude (POT1 mag. and POT3 mag.) and frequency series (POT1 freq. and POT3 freq.), and in low-flow series (Min. 7-day and Min. 30-day) for (a) station 1160650 (Mtamvuna at Gundrift), and (b) station 6545200 (Krka at Podbocje). Copyright  2005 IAHS Press

Trend detection in river flow series: 2. Flood and low-flow index series

821

Fig. 3 Trend indices (TI) for the annual maximum daily mean river flow series at 21 stations, estimated using the Mann-Kendall test and block bootstrapping. Negative trends are shown in grey dots, and positive trends in black dots, with dot sizes indicating the absolute value of the TI. The largest filled dots mark trends significant at the 10% level (|TI| ≥ 90%).

the Ann. max. series than in the POT magnitude series when a series of low peaks occur at the beginning or end of a time series with trend. These peaks may be too low to be selected for the POT analysis, whereas one per year will be included in the Ann. max. series, sometimes resulting in a steeper slope and/or longer series. For an example compare the Ann. max. and POT1 mag. time series plots for station 6545200 (Fig. 2(b)). The difference in significance between the POT and Ann. max. series suggests that it may be more useful to incorporate a larger amount of data into the trend analysis in some other way than through the simple peak-over-threshold approach used in the present study. Alternatives may include selecting a fixed number (more than one) of peaks in each year, but the characteristics of different flood indices may be better examined in a simulation study using many samples with an artificially imposed trend. An advantage of the peak-over-threshold index series remains the ability to estimate the trend in the number of floods occurring each year. A correlation analysis between the Table 2 columns for flood indices on the one hand, and the low-flow indices on the other, does not reveal anything of significance. Figure 4 and Table 2 suggest that many of the stations have experienced less severe low flows, with 10 stations showing significant positive trends for each of the Min. 7-day and Min. 30-day flow series. The increase in low flows would be consistent with an increasing number of reservoirs becoming operational in the catchments over the period of record. A reservoir’s capacity to store the incoming flood flows and slowly release the water over time generally means that low flows are augmented and flood flows are mitigated downstream of the reservoir (Vörösmarty et al., 1997). The increases in low flows are not consistent with a general intensification of the hydrological cycle in a warming climate, which would involve more severe droughts (Trenberth, 1998). However, if there is a change in the river flow regime due to reservoir construction in the catchments, this is likely to partly or completely mask changes in the hydrological cycle due to climate change. It should also be kept in mind Copyright  2005 IAHS Press

822

Cecilia Svensson et al.

Fig. 4 As in Fig. 3, but for trends in the annual minimum 7-day mean river flow series.

that any climate change impacts on river flow regimes would be a relatively recent, gradually increasing phenomenon, and that statistical tests for trend are not able to detect changes which have not lasted long, or are weak (Radziejewski & Kundzewicz, 2004a). CONCLUSIONS The investigation into trends in observed high- and low-flow index series at 21 daily mean river flow stations across the world were conducted with emphasis on the comparison of using peak-over-threshold (POT) methods as opposed to using annual maximum river flows. They suggest the following: – There is generally good agreement between the results of the trend analysis using the linear regression method and the Mann-Kendall method (significance levels for both estimated using block bootstrapping), except when there are outliers in the data series. – There is very good agreement between trends estimated for the two low-flow indices, annual minimum 7-day and 30-day mean flows. There is also good agreement between trends in the three flood magnitude series, and, separately, in the two flood frequency series, but not necessarily between flood magnitude and frequency series. – There is a larger number of both negative and positive significant trends in the annual maximum flood series, than in the peak-over-threshold magnitude series, which can be explained by the way these index series are constructed. – The results of the trend analyses are not consistent with an intensification of the hydrological cycle, as manifesting itself in an increase in floods and more severe dry spells. Rather, statistically significant increases in the low-flow series are consistent with a surmised increasing number of reservoirs becoming operational in the catchments. This imposed modification to the river flow regime would be likely to obscure any recent alteration in the hydrological cycle due to climate change. Copyright  2005 IAHS Press

Trend detection in river flow series: 2. Flood and low-flow index series

823

Acknowledgements The study was funded jointly by the World Meteorological Organization (WMO) and by the Centre for Ecology and Hydrology, UK. The study is a contribution to the WMO/UNESCO “World Climate Programme—Water” and in particular to its programme working area on “Analysing Long Time Series of Hydrological Data and Indices with Respect to Climate Variability and Change”. The daily mean river flow data were provided under a project specific licence by the Global Runoff Data Centre (GRDC, 2003), Federal Institute of Hydrology, Koblenz, Germany. The extraction of flow index series, the trend detection and the estimates of significance of trends were undertaken using the HYDROSPECT software (Radzeijewski [sic] & Kundzewicz, 2004b), upgraded for this study by Maciej Radziejewski. The use of the software, which is available free of charge at: ftp://www.wmo.int/Documents/hwr/Hydrospect.zip, is gratefully acknowledged. REFERENCES Adamowski, K. & Bocci, C. (2001) Geostatistical regional trend detection in river flow data. Hydrol. Processes 15, 3331– 3341. Burn, D. H. & Hag Elnur, M. A. (2002) Detection of hydrologic trends and variability. J. Hydrol. 255, 107–122. CEH Wallingford/Met Office (2001) To what degree can the October/November 2000 flood events be attributed to climate change? Technical report to Defra, Project FD2304, March 2001. CEH Wallingford, UK. Cluis, D. & Laberge, C. (2001) Climate change and trend detection in selected rivers within the Asia-Pacific region. Water Int. 26, 411–424. Douglas, E. M., Vogel, R. M. & Kroll, C. N. (2000) Trends in floods and low flows in the United States: impact of spatial correlation. J. Hydrol. 240, 90–105. Efron, B. (1979) Bootstrap methods: another look at the jack-knife. Ann. Statist. 7, 1–26. GRDC (Global Runoff Data Centre) (2003) Dataset of 21 river discharge time series from the Global Runoff Data Centre GRDC, Koblenz, Germany. http://grdc.bafg.de Gustard, A., Bullock, A. & Dixon, J. M. (1992) Low flow estimation in the United Kingdom. Report no. 108, Institute of Hydrology, Wallingford, UK. Hisdal, H., Stahl, K., Tallaksen, L. M. & Demuth, S. (2001) Have streamflow droughts in Europe become more severe or frequent? Int. J. Climatol. 21, 317–333. IPCC (Intergovernmental Panel on Climate Change) (2001) Climate Change 2001: Synthesis Report. Third Assessment Report of the IPCC. Cambridge University Press, Cambridge, UK. Kundzewicz, Z. W. (2004) Searching for change in hydrological data—Editorial. Hydrol. Sci. J. 49(1), 3–6. Kundzewicz, Z. W. & Robson, A. (2000) Detecting trend and other changes in hydrological data. Report WMO/TD no. 1013. World Meteorological Organization, Geneva, Switzerland. Kundzewicz, Z. W. & Robson, A. J. (2004) Change detection in hydrological records—a review of the methodology. Hydrol. Sci. J. 49(1), 7–19. Kundzewicz, Z. W., Graczyk, D., Maurer, T., Przymusińska, I., Radziejewski, M., Svensson, C. & Szwed, M. (2004) Detection of change in worldwide hydrological time series of maximum annual flow. Report to the World Meteorological Organization (WMO), World Climate Programme—Water. World Climate Programme Applications and Services (WCASP) Report WCASP–64, WMO/TD no. 1239. WMO, Geneva. Also published by the Global Runoff Data Centre (GRDC) under the same title. GRDC Report no. 32, GRDC, Koblenz. http://grdc.bafg.de/?911 Kundzewicz, Z. W., Graczyk, D., Maurer, T., Pińskwar, I., Radziejewski, M., Svensson, C. & Szwed, M. (2005) Trend detection in river flow series: 1. Annual maximum flow. Hydrol. Sci. J. 50(5), 797–810 (this issue). Kunkel, K. E., Changnon, S. A. & Angel, J. R. (1994) Climatic aspects of the 1993 upper Mississippi river basin flood. Bull. Am. Met. Soc. 75, 811–822. Lindström, G. & Bergström, S. (2003) Long-term variation in runoff and temperature in Sweden. In: Water Resources Systems—Water Availability and Global Change (Proc. Sapporo Symp., July 2003) (ed. by S. Franks, G. Blöschl, M. Kumagai, K. Musiake & D. Rosbjerg). IAHS Publ. 280, IAHS Press, Wallingford, UK. Lindström, G. & Bergström, S. (2004) Runoff trends in Sweden 1807–2002. Hydrol. Sci. J. 49(1), 69–83. Lins, H. F. & Slack, J. R. (1999) Streamflow trends in the United States. Geophys. Res. Lett. 26, 227–230. Marsh, T. J. & Bradford, R. B. (2003) The floods of August 2002 in central Europe. Weather 58, 168. Middelkoop, H., Daamen, K., Gellens D., Grabs, W., Kwadijk, J. C. J., Lang, H., Parmet, B. W. A. H., Schädler, B., Schulla, J. & Wilke, K. (2001) Impact of climate change on hydrological regimes and water resources management in the Rhine basin. Climatic Change 49, 105–128. Miller, J. R. & Russell, G. L. (1992) The impact of global warming on river runoff. J. Geophys. Res. D 97, 2757–2764. Milly, P. C. D., Wetherald, R. T., Dunne, K. A. & Delworth, T. L. (2002) Increasing risk of great floods in a changing climate. Nature 415, 514–517. Copyright  2005 IAHS Press

824

Cecilia Svensson et al.

Nijssen, B., O’Donnell, G. M., Hamlet, A. F. & Lettenmaier, D. P. (2001) Hydrologic sensitivity of global rivers to climate change. Climatic Change 50, 143–175. Radziejewski, M. & Kundzewicz, Z. W. (2004a) Detectability of changes in hydrological records. Hydrol. Sci. J. 49, 39– 51. Radzeijewski [sic], M. & Kundzewicz, Z. W. (2004b) Development, use and application of the HYDROSPECT data analysis system for the detection of changes in hydrological time series for use in WCP-Water and national hydrological services. Report to the World Meteorological Organization (WMO), World Climate Programme— Water. World Climate Programme Applications and Services Report WCASP–65, WMO/TD no. 1240. WMO, Geneva. The software and manual can be downloaded from ftp://www.wmo.int/documents/hwr/hydrospect.zip Reynard, N. S., Prudhomme, C. & Crooks, S. M. (2001) The flood characteristics of large UK rivers: Potential effects of changing climate and land use. Climatic Change 48, 343–359. Robson, A. & Reed, D. (1999) Flood Estimation Handbook, vol. 3: Statistical procedures for flood frequency estimation. Institute of Hydrology, Wallingford, UK. Robson, A. J., Jones, T. K., Reed, D. W. & Bayliss, A. C. (1998) A study of national trend and variation in UK floods. Int. J. Climatol. 18, 165–182. Saurí, D., Milego, R., Canalís, A., Ripoll, A. & Kleeschulte, S. (2003) Mapping the impacts of recent natural disasters and technological accidents in Europe. Environmental issue report no. 35, European Environment Agency, Copenhagen, Denmark. Stedinger, J. R., Vogel, R. M. & Foufoula-Georgiou, E. (1993) Frequency analysis of extreme events. In: Handbook of Hydrology (ed. by D. R. Maidment), 18.1–18.66. McGraw-Hill, New York, USA. Svensson, C., Kundzewicz, Z. W. & Maurer, T. (2004) Trends in flood and low flow hydrological time series. Report to the World Meteorological Organization (WMO), World Climate Programme—Water. World Climate Programme Applications and Services Report WCASP – 66, WMO/TD no. 1241. WMO, Geneva. Also published by the Global Runoff Data Centre (GRDC) under the title “Trends in flood and low flow series”. GRDC Report no. 33, GRDC, Koblenz. http://grdc.bafg.de/?911 Trenberth, K. E. (1998) Atmospheric moisture residence times and cycling: Implications for rainfall rates and climate change. Climatic Change 39, 667–694. Vörösmarty, C. J., Sharma, K. P., Fekete, B. M., Copeland, A. H., Holden, J., Marble, J. & Lough, J. A. (1997) The storage and aging of continental runoff in large reservoir systems of the world. Ambio 26, 210–219.

Received 24 February 2005; accepted 20 June 2005

Copyright  2005 IAHS Press