Flood drivers and trends: a case study of the Geul River catchment (the Netherlands) over the past half century

In July 2021, extreme precipitation caused devastating flooding in Germany, Belgium and the Netherlands, particularly in the Geul River catchment. Such precipitation extremes had not been previously recorded and were not expected to occur in summer. This contributed to poor flood forecasting and, hence, extensive damage. Climate change was mentioned as a potential explanation for these unprecedented events. However, before such a statement can be made, we need a better understanding of the drivers of floods in the Geul and their long-term variability, which are poorly understood and have not been recently examined. In this paper, we use an event-based approach to identify the dominant flood drivers in the Geul. We also employ (1) a multi-temporal trend analysis to investigate their temporal variability and (2) a novel methodology to detect the dominant direction of any trend. Results suggest that extreme 24 h precipitation alone is typically insufficient to cause floods. The joint probability of extreme and prolonged rainfall combined with wet initial conditions (compound event) determines the chances of flooding. Flood-producing precipitation shows a consistent increase in the winter half-year, a period in which more than 70 % of extremely high flows have historically occurred. While no consistent trend patterns are evident in the majority of precipitation and extreme flow trends in the summer half-year, an increasing direction is visible in the recent past.

1 Introduction

In July 2021, western Europe was struck by extreme precipitation, leading to disastrous flooding in Germany, Belgium and the Netherlands (Journée et al., 2023; Kreienkamp et al., 2021). This event ranked among the most devastating natural disasters to hit Europe in the past 50 years, resulting in at least 220 deaths and causing an estimated economic damage of EUR 46 billion (MunichRe, 2022; Mohr et al., 2023). In the Netherlands, the Geul River catchment (344 km²) was the most impacted, with the economic damage of the floods in this area estimated to have exceeded EUR 200 million, constituting approximately 50 % of the total estimated flood damage in the country (Task Force Fact Finding hoogwater, 2021). The event revealed weaknesses in flood risk management. Flood risks were poorly communicated to the inhabitants (Slager, 2023). The flood forecasting system for the Geul was under maintenance; however, even if it had been operational, predictions would not have been accurate (according to the responsible authorities) due to the poor representation of flood-generating processes (Task Force Fact Finding hoogwater, 2021). The Geul catchment is considered to be quite an exceptional and atypical catchment for the Netherlands due to its steep topography (elevations range from 40 to nearly 400 m) and deep soils (tens of meters). A proper understanding of flood drivers in the area is considered to be an important stepping-stone in mitigating risk in the future.

Floods in a catchment are caused by the interaction of meteorological, river system and catchment characteristics (Andrés-Doménech et al., 2015). Hydrological catchment properties can regulate streamflow response (Sharma et al., 2018), for example, extreme precipitation does not always result in floods in various basins around the world (Wasko and Nathan, 2019; Nanditha and Mishra, 2022; Berghuijs et al., 2019). Among the catchment characteristics, antecedent conditions (e.g., soil moisture) can play a crucial role in driving high flows (Bertola et al., 2020; Woldemeskel and Sharma, 2016). Many locations around the world have experienced the effects of wet antecedent conditions on flood risk (e.g., Garg and Mishra, 2019; Bischiniotis et al., 2018; Ivancic and Shaw, 2015; Cao et al., 2019). Especially in lowland catchments, the discharge response can be strongly influenced by the catchment wetness, due to shallow groundwater and its effects on rainfall flow paths (Brauer et al., 2018). As a result, determining the relative contributions of antecedent wetness conditions and extreme precipitation to high river flows is critical.

The identification of the drivers of observed flood events has received increasing attention in the recent literature (e.g., Blöschl et al., 2019; Bertola et al., 2020). Examining relations between trends or seasonality in flood peaks and factors such as extreme precipitation or soil moisture to define flood drivers is well established (Blöschl et al., 2019; Do et al., 2017; Tramblay et al., 2021; Wasko et al., 2020). However, to enhance our knowledge of flood dynamics, an event-based approach has been suggested (Nanditha and Mishra, 2022; Tramblay et al., 2021; Berghuijs et al., 2019). This approach entails identifying the specific drivers behind individual flood events or extremely high flow occurrences. By analyzing the characteristics and circumstances surrounding these events, one can gain valuable insights into the mechanisms and factors that contribute to their intensity and occurrence (Nanditha and Mishra, 2022).

A proper understanding of flood drivers not only includes their identification but also their long-term change. Investigating the changes in variables that can cause a hazard is crucial for managing risk in an effective way (Yang et al., 2021) and can facilitate the planning of reliable and meaningful interventions. Making a critical assessment of the past and current states and providing the long-term trends in hydroclimatic variables play a key role in future projections (Squintu et al., 2021). Trends and changes in the time series of hydrological and climatological data have received attention in catchments and areas across the world (e.g., Blöschl et al., 2017; Do et al., 2017; Hannaford et al., 2021; Murphy et al., 2020).

However, most of the existing trend tests are limited, as they are conducted within fixed time frames, which may fail to accurately capture historical variability. The significance and magnitude of trends can vary considerably based on the chosen study period and duration. To deal with this limitation, multi-temporal trend approaches have been leveraged (e.g., Hannaford et al., 2021; Murphy et al., 2020), considering all possible combinations of start- and end-year periods. Although these analyses have helped to identify temporal variability, a research gap remains in determining the main trend direction, such as consistency or stability, across all studied time frames. Łupikasza (2010) developed criteria for expressing trend stabilities using a fixed 30-year moving window; however, this approach did not fully consider all of the variability (multi-temporal analysis) and, instead, utilized an overlapping period, which can be misleading. As each 30-year window overlaps with the previous and subsequent windows, trends can be missed or misinterpreted because the overlapping periods can obscure them. Thus, using only overlapping periods can lead to artificially smooth trends that do not represent the true (long-term) variability in the data. In addition, the length of the selected moving window might introduce bias in the analysis. Trends over longer periods and with different combinations of start- and end-year periods expressing the full historical variation should also be taken into account in the calculation of the main direction. To address these limitations, our study builds on the multi-temporal approach and develops a methodology capable of identifying and assessing trend consistency in multi-temporal analyses, considering the complete range of variability. This new method is anticipated to deepen our understanding of flood driver trends in the Geul River catchment, with potential applicability across broader contexts.

In summary, the Geul River catchment is an interesting example of a hilly catchment in northwestern Europe, with unique hydrological characteristics that are shaped by the underlying geology, topography and land use. The recent floods have shown that there is a need for further research into the drivers of flooding in the area, particularly in the context of climate change. Specifically, the role of extreme precipitation and antecedent conditions as potential flood drivers and their long-term variability remain to be examined. Therefore, our objective is to detect the primary drivers of high-flow/flood events in the Geul River catchment and analyze their long-term trends. To achieve these objectives, we address the following scientific questions that are crucial for our understanding of floods in the Geul River catchment:

• a.

  What are the dominant flood drivers in the Geul River catchment in different seasons?

• b.

  What are the trends (temporal variability and consistency) in the flood-producing precipitation and extreme discharges in the catchment?

To identify the dominant flood drivers, we use an event-based approach. In addition, we utilize a multi-temporal trend analysis to investigate the temporal variabilities in the trends and introduce a new methodology to detect the dominant direction (i.e., consistency) of a trend. Although our study focuses on the Geul area, it is essential to highlight that our combined approaches (integrating an event-based approach with multi-temporal analyses) and proposed trend consistency method are applicable beyond this specific case. Thus, our aim is to offer valuable insights for the Geul area while avoiding constraining the scope of our methods and findings to a singular case study.

Figure 1 (a) Location of the Geul River catchment (shaded blue area) and (b) elevation map of the study catchment, including the location of precipitation stations and the discharge observation station at the outlet.

Figure 2 The upper subpanel of panel (a) shows the monthly averages for precipitation (averaged using Thiessen polygons), reference (Makkink) evaporation (obtained from the Maastricht meteorological station) and discharge in the Geul catchment; the lower subpanel presents flow regimes of the Geul based on the Pardé coefficient, defined as the ratio of the average monthly discharge to the average annual discharge. Panel (b) shows a flow duration curve along the Geul based on the mean daily discharge. All graphs are produced using time series from 1970 to 2021.

2 Methods and materials

2.1 Study area

The Geul River is an important tributary of the Meuse and is located in the Netherlands, Belgium and part of Germany, close to the three-border region (Fig. 1). The total area of the Geul catchment is approximately 344 km². The Geul drops about 250 m over approximately 60 km, making it one of the few steeply sloping rivers in the Netherlands. The Geul River has an average discharge of approximately 3.2 m³ s⁻¹ (at the outlet of the catchment) and is mainly rain-fed. As a consequence, its discharge can change dramatically during flood and drought events (e.g., ranging from 1 m³ s⁻¹ during drought periods to more than 40 m³ s⁻¹ during floods). The response time of the catchment is in the order of 1–2 d (Asselman et al., 2022). The annual average precipitation is approximately 870 mm yr⁻¹ and is rather uniformly distributed over the year (see Fig. 2a). Average annual discharge at the outlet of the catchment and potential evaporation are about 307 and 585 mm yr⁻¹, respectively, based on time series from 1970 to 2021. The flow regimes in the Geul do not show large variations (Fig. 2). On average, the equal distribution of runoff in the Geul throughout the year is due to the effect of groundwater storage provided by large chalk aquifers in the catchment (Tu, 2006). The long-term evaporative index is approximately 0.67 and the runoff ratio is 0.35.

2.2 Data sets

Streamflow time series at the outlet of the catchment (Meerssen station), were made available by the local water authority, Waterschap Limburg, including 15 min measurements from 1970 up to August 2021. From 1970 to 2011, measurements were taken using a measuring weir; however, since 2011, discharges have been measured using acoustic Doppler current profilers (ADCPs) (van der Deijl, 2023). Historical flood event data (i.e., date of occurrence) are based on Thewissen (2022), who performed a local newspaper search using Delpher, a Dutch database containing digitized texts from newspapers, books and magazines. These collections are curated by scientific organizations, libraries and heritage institutions. Thewissen (2022) obtained the data through an iterative process that involved optical character recognition (OCR) and manual scanning. In this research, a flood is defined as surpassing the bankfull capacity. For detailed information about the search methodology, please refer to Thewissen (2022).

Records of 24 h precipitation from five stations located in (or near) the Geul River catchment are used (see Fig. 1). The data used in this study come from the Royal Netherlands Meteorological Institute (KNMI) manual rain gauge network. Volunteer observers operate the rain gauges on a 24 h basis (daily, from 08:00 until 08:00 UTC the following day). In addition to the volunteer KNMI stations, daily (calendar day) measurements from the automated meteorological station at Maastricht are used. Data from automated stations are also available on an hourly basis; however, the daily scale is used in order to have the same resolution for each station. The time series are complete (except for 3 missing months at Noorbeek) and are considered to be of high quality, as KNMI performs regular quality tests (Buishand et al., 2013). Days with extremes were checked, and it was found that all of them occurred during well-known high-precipitation or flood events.

The available precipitation record periods slightly differ between the selected stations and mainly start in the 1950s, except the new Epen station, which has a record from 1981 onward. Specifically, stations at Vaals, Valkenburg and Noorbeek have a record from 1951 onward, while the available time series at the Maastricht and Ubachsberg stations start from 1958 and 1955, respectively. Due to its shorter available period, Epen station is excluded from the analysis. The precipitation stations located in the uppermost regions of the catchment in Belgium were not taken into account due to their limited temporal coverage, lack of alignment with the discharge (and KNMI) data and unsuitability for trend analysis.

Another important factor to consider when analyzing meteorological time series, and especially trends, is data (in)homogeneity. Trends represent the long-term changes in the data over time caused by natural/climate variability, while inhomogeneities represent changes in the data that are not part of the underlying trend. There are several types of modifications that can occur and can cause significant non-climatic alternations (inhomogeneities) in the data, especially in long time series, such as changes in the location of the measuring station, differences in the manner and the procedures (e.g., measuring frequency), or changes in the instruments or tools. Performing statistical tests for homogeneity on precipitation data measured at daily frequencies is challenging or practically infeasible (Łupikasza, 2010). This can be attributed to the fact that daily time series show strong random variations. For this reason, the precipitation time series were aggregated to monthly sums and then tested using two well-known homogeneity tests: the standard normal homogeneity test (SNHT) (Alexandersson, 1986; Alexandersson and Moberg, 1997) and Buishand's U test (Buishand, 1984). Both tests were applied at a significance level of 5 % (α = 0.05). The test results showed that the data time series were free of significant errors and that no inhomogeneities were detected.

2.3 Flood driver detection

2.3.1 Extreme indicators

Hydrological years are used instead of calendar years to ensure a more accurate representation of the frequency and intensity of extreme events that may occur within a particular hydrological regime. Similarly to Tu (2006) and Sperna-Weiland et al. (2015), we define a hydrological year from November to October, for the study area. Half-year hydrological winter (from November to April) and summer (from May to October) are also considered.

Two types of flood data time series are used: annual maxima (denoted as Q_max), including winter and summer yearly maxima (denoted as Q_W,max and Q_S,max, respectively), and extremely high flows using the peaks-over-threshold (POT) method (Haan, 2002). The maximum daily discharge is taken from the 15 min discharge data, and hydrological years with over 20 % missing daily values are omitted from the analysis. Consequently, the years 1971, 1974 and 1990 have been excluded from the Q_max times series. For identifying extremely high flows, we use the 99th percentile threshold (denoted as Q₉₉; Nanditha and Mishra, 2022) extracted from the maximum daily discharge time series from 1970 to 2021, excluding daily missing values. We use only extremely high events separated by a time frame of 5 d to ensure that the selected high flows are independent and do not belong to the same flood wave.

Floods are generally caused by a combination of initial moisture conditions and precipitation. To explore the probable causes of high-flow episodes, we employ six indicators to assess precipitation and antecedent soil moisture levels: extreme precipitation (denoted as P₉₉), multiday precipitation (denoted as P_MD), wet antecedent conditions (denoted as P_WAC), P₉₉ combined with P_WAC (denoted as Compound I), P_MD combined with P_WAC (denoted as Compound II), and P₉₉ and P_MD combined with P_WAC (denoted as Compound III). These indicators allow us to examine the relative role of extreme precipitation; prolonged, heavy rainfall; extreme soil moisture conditions; and compound extremes in generating high flows.

We estimate P₉₉ as the events that exceed the 99th percentile of wet days (days with more than 1 mm of precipitation) (Nanditha and Mishra, 2022). We define P_MD events using the 95th percentile of all k-day accumulated (rolling sum) precipitation time series (Nanditha and Mishra, 2022). To clarify the P_MD definition, we ensure that the 95th percentile of multiday rainfall consistently surpasses the 99th percentile of the 24 h rainfall on wet days, aiding in distinguishing between P₉₉ and P_MD. In this way, usually more than 2 d of precipitation are necessary to exceed the k-day 95th percentile and trigger P_MD, allowing the assumption that P_MD can be used as a proxy for heavy, prolonged events. As we use the 95th percentile of all k-day accumulated (rolling sum) precipitation to define P_MD and employ “daily” values, this threshold is expected to be exceeded in prolonged events irrespective of the selected duration, indicating that we have prolonged (multiday) heavy events (larger than the 95th percentile of the selected k-day accumulations), although not as extreme as the 24 h P₉₉, which helps us examine the relative contributions of extreme precipitation and prolonged, heavy rainfall in generating high flows. However, in rare cases, 24 h precipitation can simultaneously trigger both P₉₉ and P_MD especially for the lower k-day accumulation periods, which is unavoidable. Thus, for each of the five precipitation stations considered, we calculated the P_MD 95th percentile for different durations. It was found that a duration longer than 4 d is required for this percentile to surpass the 99th percentile used in defining P₉₉. Finally, to determine the most suitable k-day P_MD duration for k ≥ 4, we evaluate the P_MD probability preceding high flows across multiday precipitation durations of up to 10 d (see Sect. 3.1.1).

Furthermore, we use the antecedent precipitation index (API; the ratio of the 30 d pre-event precipitation to the long-term average for the same period), as developed by Marchi et al. (2010), to assess the initial catchment wetness and get an estimate of the initial (soil) conditions. Marchi et al. (2010) classifies antecedent moisture conditions as follows: (1) dry, 0 < API ≤ 0.5; (2) normal, 0.5 < API ≤ 1.5; and (3) wet, API > 1.5. P_WAC corresponds to API values higher than 1.5 (Marchi et al., 2010). The API's effectiveness in assessing initial soil wetness conditions has been documented, for instance, by Marchi et al. (2010), who demonstrated its strong agreement with predictions from a continuous soil moisture accounting hydrological model (Norbiato et al., 2008). However, as the index is based solely on precipitation, its sensitivity to evaporation is further discussed in Sect. 3.1.5. This is done by computing the 30 d pre-event effective rainfall, which entails subtracting reference evaporation obtained from the Maastricht station from the precipitation measurements.

2.3.2 Monthly distribution of extremes

In order to get a rough indication of the effects of extreme rainfall on high flows, we estimate the monthly distribution of P₉₉, P_MD, annual maximum precipitation (denoted as P_max) and high flow extremes (i.e., Q₉₉ and Q_max) as well as past flood events (as defined in Sect. 2.2). Despite the fact that the monthly distribution of P_WAC events cannot be directly related to the monthly distribution of high flow extremes, as P_WAC events indicate wetter-than-average conditions any time of the year, their relative frequencies can provide useful insights. For this reason, we also calculate the monthly frequencies of P_WAC events. To achieve this, each daily time step in the precipitation series is treated as an individual event. This involves summing the precipitation amount for each day over the previous 30 d and then dividing this sum by the 30 d long-term average for the same period across the entire time series. In this way, an API index per day is obtained.

The use of all-day percentiles for discharge (Q₉₉) and wet-day percentiles for rainfall (P₉₉) may lack statistical robustness, primarily because there could be a potential increase in the number of wet days over the study period (Schär et al., 2016). To address the potential issue, a trend test with respect to the number of wet days per year and station was conducted using the nonparametric Mann–Kendall (M–K) (Kendall, 1955; Mann, 1945) test at a significance level of 0.05. The analysis did not reveal any significant trends in the number of wet days.

Table 1 Extreme indicators used to identify the main flood drivers. The initial six indicators are used to calculate probabilities preceding high-flow episodes, whereas the last indicator is correlated with high flows.

2.3.3 Event-based analysis

After the seasonal assessment of the extreme indicators and high-flow events, an event-based approach is followed to detect the predominant flood driver(s) in the catchment. Specifically, we calculate the likelihood of the six precipitation indicators (i.e., P₉₉, P_MD, P_WAC, Compound I, Compound II and Compound III) preceding selected high-flow episodes (i.e., Q_max) in order to determine the primary flood cause. We use the Q_max events instead of the Q₉₉, as they are better defined for identifying flood drivers including all major flood events, also considering that the P₉₉ is calculated from a significantly longer data series than the Q₉₉. We conduct a comprehensive analysis of each high-flow event, examining the frequency of occurrence for each indicator before a Q_max event. Specifically, we determine how many times an indicator occurred in relation to the total number of Q_max events, effectively calculating its likelihood of contributing to such occurrences. For instance, if P₉₉ is the main flood cause in the catchment, then P₉₉ is highly likely to occur during the duration of the flood (precipitation amount on the same day) or within 1 d before the commencement of the event. Here, we use a 2 d interval (i.e., precipitation on the same or previous day of the flood event) for two main reasons: (1) the lag time of the catchment and (2) to eliminate errors due to the use of daily values (e.g., a high-flow event may occur early in the morning due to P₉₉ from the previous day). For Compound-I events (P₉₉ under P_WAC), we verify whether the P₉₉ under P_WAC occurred either on the event day itself or 1 d prior to the event. This approach guarantees that the P₉₉ consistently occurs under preexisting wet conditions. Thus, we establish the requirement for a P₉₉ to appear under P_WAC for a compound event, thereby preventing scenarios in which a P₉₉ occurring 1 d prior to the Q_max under normal circumstances could increase the API on the day of the Q_max, leading to P_WAC. In a similar manner, we calculate the probability of Compound III preceding the Q_max events: P_MD and P₉₉ under P_WAC on the day of the event or 1 d before. For the remaining indices, P_MD and P_WAC, we simply verify whether these indicators are present on the day of the event.

As we directly compare the likelihood of each indicator leading to Q_max, it is important to note that the comparison among different flood drivers may be affected by the rarity of certain events. The total number of flood indicators varies, particularly for the most extreme ones, such as P₉₉, Compound I and Compound III. For instance, in some years, a P₉₉ event might not occur; therefore, it cannot precede Q_max. To account for this effect, the reverse scenario is also considered: given that a driver has occurred, what is the relative frequency of it being followed by a Q_max event. Thus, we calculate the number of separate hydrological years in which a driver is observed and determine how often Q_max is also observed within these instances. Additionally, we provide general information about the frequency of the different drivers occurring, regardless of whether Q_max has occurred, along with their corresponding discharge values. In this reverse extreme-precipitation-based analysis, the highest discharge is extracted either on the day of the indicator or the next day for P₉₉, Compound I and Compound III, while the discharge value on the day of the event is taken for other indices to ensure alignment with the extreme-discharge-event-based analysis.

Antecedent catchment wetness and precipitation depths are crucial factors for floods in terms of magnitudes and volumes. In addition to the aforementioned flood indicators, we delve deeper into this relationship by employing the nonparametric Spearman rank-order correlation method to correlate the half-year maximum discharge time series (Q_W,max and Q_S,max) with their k-day antecedent precipitation depth (denoted as P_kD, where k = 1, 3, 5, 7, 10, 15, 30 and 40 d; Tu, 2006). The antecedent precipitation depths are computed by summing the precipitation on the day of the high-flow event and the previous k days (Tu, 2006). Table 1 provides an overview of all of the different flood indicators used in this study as well as their corresponding definitions.

One crucial concern is the timing difference in the measured daily values, i.e., calendar days for discharges and 24 h sums from 08:00 UTC for precipitation. The reported time/date in the precipitation time series (for the manual rain gauge network; Valkenburg, Ubachsberg, Noorbeek and Vaals stations) is the end time of observation. For this reason, the time of occurrence of the max daily discharge is checked: if the time of the maximum 15 min discharge values is observed to be between 00:00 and 08:00 UTC, the calendar day of this event is reported; otherwise, the next day is considered to be the date of occurrence of the event for the (manual rain gauge) precipitation time series. In this way, we ensure the agreement between the precipitation and discharge time series.

2.4 Trend analysis

The temporal variability and the trends in the aforementioned potential flood drivers (i.e., P₉₉, P_MD and P_kD) are investigated. Trends in P_kD are based on the highest k-day total precipitation per year (a summation moving window with different lengths is applied over the whole time series from the 1950s to 2021, and the annual maxima are extracted). In contrast to the definition of P₉₉ in Sect. 2.3.1, we use the 95th percentile, as events that exceed the 99th percentile are extremely rare and result in many zero values in the time series, especially in winter periods, making the trend analysis unstable. The new index is denoted as P₉₅ and represents the annual total precipitation from days exceeding the 95th percentile. Similarly to P₉₉, the 95th percentile is calculated using the whole range of the time series and only wet days. Finally, the P_MD trends are investigated by summing the annual amount of 4 d accumulated precipitation that exceeds the 95th percentile thresholds (of the 4 d rolling sum). The used precipitation trend indices are similar to those defined by ETCCDI (Expert Team on Climate Change Detection and Indices) and have been frequently applied (Klein Tank et al., 2002; Dunn et al., 2020). The aforementioned precipitation indices are calculated for winter and summer periods. In addition, trends in the half-year maximum discharges (Q_W,max and Q_S,max) are examined in order to detect possible consequences of extreme precipitation on extreme streamflows. We use the nonparametric M–K test to detect significant trends. To this end, a Python package that contains all of the types/modifications of the M–K test, as developed by Hussain and Mahmud (2019), is used. The original M–K test is employed on both the precipitation indices and discharge time series, instead of a modified M–K version that accounts for the influence of serial correlation on trend calculations. This choice is guided by the assumption that the precipitation time series exhibit no significant serial correlation and that the annual maximum discharge values are typically considered to be uncorrelated by construction.

The statistical significance and direction of the trends in the multi-temporal approach are used in order to determine the temporal consistency of a trend for each precipitation index. We express consistency as the percentage of time (t) during which trends are statistically significant. Trends that are significant at α = 0.2 (Łupikasza et al., 2011) are defined as statistically significant trends. As precipitation is characterized by strong temporal and spatial variation, the statistical significance levels can be lower compared with other climatic variables (Łupikasza et al., 2011). In addition, the large number of calculated trends (more than 850 for most stations) allows the use of lower significance levels for expressing stabilities or consistencies. In a multi-temporal approach, it is preferable to focus on the direction and intensity of the trends, rather than whether they surpass a strictly predetermined and somewhat arbitrary level of significance (Hannaford et al., 2013). A trend in an index is considered to be inconsistent, weakly consistent, consistent or strongly consistent according to the following criteria: (1) inconsistent, 0 % < t ≤ 15 % or the numbers of significant increasing and decreasing trends are similar (i.e., the percentage of significant trends in the same direction ranges between 40 % and 60 %); (2) weakly consistent, 15 % < t ≤ 25 % and more than 60 % of significant trends are in the same direction; (3) consistent, 25 % < t ≤ 45 % and more than 60 % of significant trends are in the same direction; and (4) strongly consistent, t > 45 % and more than 60 % of significant trends are in the same direction.

3 Results

3.1 Identification of the dominant flood driver

3.1.1 Selection of the P_MD duration

In this section, we explore the sensitivity of the probability of P_MD (95th percentile of all k-day accumulated time series; see Sect. 2.3.1) occurrences before high flows (i.e., Q_max) with respect to the duration of multiday precipitation. Our analysis reveals that a 4 d duration is the most suitable for defining P_MD, as the average relative frequency of Q_max preceded by this duration is the highest compared with other accumulation periods (i.e., P_MD durations of 5, 6, 7, 8, 9 and 10 d; Table 2). This finding aligns with the hydrological behavior of the catchment, as documented by Asselman et al. (2022). Additionally, we observe that the average relative probability of P_MD decreases by approximately 10 % when the duration increases from 4 to 5 d. However, for durations longer than 5 d, the frequency of P_MD preceding high flows remains relatively stable. In light of these results, the 4 d duration is utilized as the standard for P_MD definition.

Table 2 Mean relative frequencies (for all stations) of high-flow events preceded by multiday precipitation (P_MD) for different accumulation periods.

Figure 3 Monthly distribution of annual maximum precipitation (P_max), extreme precipitation (P₉₉), multiday precipitation (P_MD) and wet antecedent condition events (P_WAC) at the Vaals station as well as annual maximum discharge (Q_max), extremely high flows (Q₉₉) and past flood events in the Geul.

3.1.2 Seasonality of extreme indicators

The seasonal distributions of extreme precipitation, high-flow events and flood drivers indicate that extreme precipitation, although more frequent during summer months, does not consistently coincide with high-flow events (Fig. 3). Specifically, Fig. 3 shows the seasonal distribution of extreme precipitation (P₉₉; 69 events – about 0.4 % of the daily time series from 1970 to 2021), annual maximum precipitation (P_max; 52 events), multiday precipitation (P_MD; 933 events – 5 % of the 4 d accumulated time series from 1970 to 2021) and wet antecedent condition events (P_WAC; 2705 events – 14 % of the 30 d accumulated time series from 1970 to 2021) at the Vaals station, as an example, as well as high-flow events (Q_max and Q₉₉; 49 and 91 events, respectively) and past flood events (30 events) in the catchment. General information on the frequency of the flood drivers, averaged for all rainfall stations, is provided in Table 5. An opposing seasonality is visible between extreme precipitation and high-flow events. The relative frequencies of P₉₉ and P_max in half-year summer periods are 75 % and 73 %, respectively, while these percentages are only 22 % and 29 %, for the Q_max and Q₉₉ events, respectively. This pattern is also verified by the past flood events: only 26 % of them occurred during half-year summers. P_MD appears to occur with relatively similar frequencies throughout the year. Similar to P_MD, P_WAC shows an equal distribution over all seasons, as expected. As precipitation is rather uniformly distributed all over the year (see Fig. 2a), wetter-than-average conditions can occur in any season, and the monthly distribution of P_WAC cannot be directly linked to the seasonal distribution of high flows. However, wetter conditions (indicated by higher API values) are expected to have different effects on high flows, and a closer examination of the specific condition before every high-flow event is necessary. Overall, our findings indicate that extreme 24 h precipitation is not the most critical driver of high flows. Although extreme precipitation events tend to occur more frequently during the summer months, high-flow and flood events do not align with these periods. Factors such as antecedent soil moisture conditions and the timing, duration and intensity of rainfall events may exert a more significant influence on high-flow generation in the catchment. Therefore, greater attention is required with respect to understanding these factors.

Figure 4 Relative frequencies of Q_max being preceded by (a) extreme precipitation (P₉₉), (b) multiday precipitation (P_MD), (c) wet antecedent conditions (P_WAC), (d) extreme precipitation under wet antecedent conditions (Compound I), (e) multiday precipitation under wet antecedent conditions (Compound II) and (f) extreme and multiday precipitation under wet antecedent conditions (Compound III), with respect to the total Q_max events (count of a driver leading to Q_max in the Q_max cases divided by the total number of cases). The mean relative frequency (in %) for all rainfall stations is reported in brackets. The mean conditional probabilities of P₉₉ and/or P_MD preceding Q_max given that P_WAC precedes Q_max are also reported per compound indicator.

3.1.3 Extreme-discharge-event-based analysis

Figure 4 illustrates the relative frequencies of the introduced indicators (see Table 1) preceding high-flow episodes (i.e., Q_max), as described in Sect. 2.3.3. In approximately 75 % of the Q_max cases, a P_MD precedes high-flow events (Fig. 4b), while the corresponding percentage for P_WAC is approximately 48 % (Fig. 4c). In most cases, P_WAC should be followed by P_MD in order to generate high flows (similar percentages between P_WAC, as seen in Fig. 4c, and Compound II, as seen in Fig. 4e). This is also visible in the calculation of the conditional probability of P_MD preceding Q_max given that P_WAC precedes Q_max, which is approximately 84 % (Fig. 4e). In other words, given that P_WAC precedes Q_max, there is an 84 % chance that P_MD will be followed by Q_max. The percentages between Compound-I and Compound-III events (Fig. 4d and f, respectively) are the same, showing that each time a Q_max is preceded by P₉₉, it also preceded by P_MD. This effect, in most cases, is caused by the high amount of precipitation that fell on the day of the event or the previous day (definition of P₉₉), which also increases the 4 d precipitation (higher than the P_MD 95th percentile).

Figure 5 Annual maxima events (Q_max) and their antecedent precipitation index (API) values at Maastricht (a) and Vaals (d), including their preceding extreme indicators. Orange markers denote events preceded solely by P_MD, green markers indicate events preceded exclusively by P₉₉, purple markers represent events preceded by both P₉₉ and P_MD (thus classified as both P₉₉ and P_MD in Fig. 4), and blue markers signify events without any extreme precipitation indicator preceding them. Purple and orange markers within the “Wet” classification, along with P_WAC, P₉₉ and/or P_MD classifications, are also classified as Compound-II events in Fig. 4. Discharge events preceded by Compound III (and thus Compound I and II) are indicated with red circles. The top five floods during the study period are shown with their year of occurrence. The dashed purple line represents the linear fit, using the least-squares approach, between the API of the high-flow events preceded by P₉₉ and P_MD and their respective Q_max values, while the red dashed line represents the linear fit between the API of the Compound-III events and their Q_max. The total 4 d precipitation versus Q_max is presented for these events at Maastricht (b) and Vaals (e); the highest 24 h precipitation (highest of the two P₉₉ amounts on the day of the event or the previous day) versus Q_max at Maastricht (c) and Vaals (f) is also shown. The shaded area shows the 95 % confidence intervals for the fits, and the Pearson correlation coefficients (ρ) are also reported.

Figure 5 shows the Q_max events plotted against their API values, including their preceding precipitation indicators (i.e., P₉₉ and P_MD) at the Maastricht and Vaals stations. This figure actually presents how the different events are classified based on the previously defined indicators (Table 1), emphasizing the influence of wet conditions on high flows and exploring correlations between Q_max and associated precipitation amounts (P₉₉ or P_MD). For example, all P_MD markers (both orange and purple markers), irrespective of their wetness (API), are classified as P_MD and, thus, used to calculate the relative frequencies of Q_max being preceded by P_MD in Fig. 4b. Furthermore, Fig. 5 reveals overlapping event classifications, where one event can align with multiple indicators at the same time, e.g., a Compound-III event is classified as P_WAC, P₉₉, P_MD, Compound I and Compound II, while a P₉₉ event can concurrently be P_MD. A Q_max event preceded by P₉₉ may appear in the “Wet” classification without being classified as Compound I. This is because we specify that P₉₉ should occur under existing P_WAC to be classified as Compound I. This condition is imposed to prevent a P₉₉ event from inflating the API the day before the event, potentially leading to an API > 1.5 on the day of the event (see Sect. 2.3). For example, the 1970 event at Vaals (Fig. 5d), which is preceded by P₉₉, P_MD, P_WAC and Compound II, is not classified as Compound I or Compound III.

Table 3 Top five floods that occurred in the catchment during the study period, their rainfall accumulations and an estimate of their initial conditions based on the API. The day with the highest recorded discharge is considered to be the date of occurrence of the event. The different precipitation sums (and subsequently API values) are based on daily precipitation records at the Maastricht (first value) and Vaals (second value) stations.

All high-flow events appear to happen under normal or wet conditions (Fig. 5). Most extreme events are caused by a combination of P₉₉ and P_MD or just P_MD. Antecedent conditions also play a crucial role in translating precipitation extremes to high flows. We observe that higher API values (higher wet initial conditions) lead to higher peak discharges, especially for events preceded by compound and P₉₉ events, where a strong correlation between Q_max and the API is observed, while a very weak correlation between Q_max and the total event P_MD or P₉₉ precipitation is reported (see Fig. 5). P_WAC makes the difference between a “regular high flow” and a flood event. This is also evident in the initial conditions of the top five floods that occurred in the catchment during the study period (see Fig. 5 and Table 3). Details about these events (i.e., date of occurrence, different rainfall accumulations and an estimate of their initial conditions) are presented in Table 3 based on the precipitation records at Maastricht and Vaals. Examining the preceding conditions for the major past floods, it appears that, in most of these cases, although the precipitation events spanning 1–3 d were heavy, the overall precipitation over the 30 d preceding the events was also substantial. This extended period of precipitation likely played a critical role in saturating the catchment, making it more susceptible to flooding. The combination of intense rainfall over shorter durations and continuous precipitation over the 30 d period seemed to collectively contribute to the formation of wet initial conditions, ultimately increasing the risk and eventually resulting in flooding. It is also important to note that the large discrepancies in rainfall accumulations between the two stations found in Table 3 are caused mainly by the different time intervals that they include, i.e., calendar days for the Maastricht station and 24 h sums from 08:00 UTC for the Vaals station, as well as the spatial variability.

The absence of green markers in Fig. 5 (only P₉₉ events) indicates again that, for days preceding Q_max, the probability of a P_MD event given that P₉₉ occurs equals 1 (Pr(P_MD|(P₉₉ and Q_max)) = 1). In the way that we defined P₉₉ (2 d interval), we observe that the P₉₉ events preceding a Q_max usually also coincide with P_MD. For very extreme 24 h events, the 4 d 95th percentile used for the P_MD definition can be exceeded and concurrently cause both P_MD and P₉₉, which is unavoidable. However, in longer accumulation periods for P_MD (i.e., 5, 6, 7, 8, 9 and 10 d), the corresponding 95th percentile increases, as the moving/accumulated period is extended and becomes much larger than the 99th percentile used for the definition of P₉₉. In these cases, irrespective of the duration, the mean relative frequencies of high flows preceded by Compound II and Compound III remain stable (see the Supplement for the analysis). This implies that rainfall events (whether heavy or not) probably preceded P₉₉ for these events as well (at least for less extreme ones), potentially resulting in wet conditions and consequently high discharges, highlighting the correlation among the different drivers used and how they can be converted to compounds. Thus, while it is found that Q_max is preceded by P_MD 75 % of the time, some of the P_MD events could be forced or even caused by P₉₉. However, the definition of P_MD still holds significance, as it denotes an extended period of heavy rainfall.

Table 4 Correlation coefficients between the winter half-year discharge maxima and their antecedent k-day precipitation depths.

Table 5 General information, averaged for all rainfall stations, regarding the frequency of the different drivers occurring, regardless of Q_max having occurred, in the Q_max period (hydrological years from 1970 to 2021, excluding 1971, 1974 and 1990).

Finally, we also explore the effect of antecedent precipitation depths on high flows, as an indicator of antecedent catchment wetness. Table 4 reports the correlation coefficients of the Q_W,max time series with their k-day (k = 1, 3, 5, 7, 10, 15, 30 and 40) antecedent depths per rainfall station. Results are only presented for the winter half-year, as approximately 80 % of Q_max occurred in this period (Fig. 3). Thus, Q_S,max time series contains mainly low discharge values (most of them lower than 20 m³ s⁻¹; see Fig. 10a) with meaningless correlations. Peak half-year winter discharges in the Geul appear to be closely related to antecedent 10–40 d precipitation depths on wet soils (e.g., correlation coefficient of 0.66 in Vaals for the duration of 15 d). The effect of wet antecedent conditions is also reflected here. These results seem reasonable considering the opposing seasonality observed in the catchment (Fig. 3). The July 2021 flood event is consistent with these results, as there is evidence that the wetness of the catchments was much higher than usual for the time of the year (see Fig. 5 and Table 3).

3.1.4 Extreme-precipitation-based analysis

In this section, we approach the problem from an inverse perspective compared with Sect. 3.1.3. Table 5 shows the frequency of the potential flood drivers actually leading to extreme discharge. P₉₉, Compound I and Compound III are rare events compared with the other drivers, and there are years in which they are not observed at all (see Table 5). In the separate years when a P₉₉ event occurs, it is followed by a Q_max event in about 37 % of those cases, while these numbers are approximately 65 % for Compound I and Compound III, respectively (Table 5). Compound-I and Compound-III events are extremely rare events (approximately 13 recorded events per station); however, in most cases, when they appear, they lead to Q_max. This observation suggests that when extreme precipitation P₉₉ occurs under wet antecedent conditions, it leads to high flows; however, alone, it is typically insufficient to cause floods.

Figure 6 Empirical cumulative distribution functions (eCDFs) of discharge caused by the extreme precipitation indicators for all stations, based on the number of events from Table 5.

It appears that flooding is rarely caused by a single mechanism. Figure 6 shows the discharge empirical cumulative distribution functions for every station caused by every extreme precipitation indicator, for the number of events reported in Table 5. According to the local water authorities, a discharge exceeding approximately 40 m³ s⁻¹ could lead to flooding of the first houses, while a discharge greater than 30 m³ s⁻¹ could result in the inundation of floodplains, particularly concerning upstream locations within the catchment area (Klein et al., 2024). These thresholds are often crossed during Compound-I and Compound-III events, particularly in extreme events at the Vaals station, indicating a stronger correlation with discharges compared with other locations. It is important to note that these discharge thresholds pertain to the upstream sections of the catchment, not the outlet at Meerssen. However, most of the water originates from the Belgian part, while the contribution of the remaining tributaries is minor (Klein, 2022). Overall, it seems that single processes are not likely to cause floods, whereas compound events do. Moreover, it is again clear that P₉₉ in isolation is much less likely to cause floods, compared with when it is combined with wet antecedent conditions and P_MD.

3.1.5 Sensitivity of the API to evaporation

The exclusion of evaporative processes in the API used for evaluating initial catchment conditions could potentially pose significant concerns, as the index is solely based on antecedent precipitation depths. An important aspect to consider is whether the API reliably reflects soil wetness consistently throughout the year in our study region or if its interpretation is influenced by the strong seasonal variations (reference evaporation is low in winter and high in summer; see Fig. 2). To investigate this, the effective rainfall (precipitation minus reference evaporation from the Maastricht station) 30 d before an event is calculated instead of the API, and Fig. 5 is reproduced.

The results suggest that no significant biases occurred due to the exclusion of evaporation. Larger offsets are visible in the summer half-year events and the overall correlation is low (as in the regular API; Fig. 5); however, the high correlation between the P₉₉ and P_MD events and the Q_max (purple markers and reported correlation value) is maintained (Fig. 7), which is in line with the API results. Only 7 of 49 Q_max events occurred in summer, and 6 of these 7 summer Q_max events are in the P₉₉ and P_MD events; as a result, the calculated correlation includes them. In addition, the top five floods remain “higher” compared with other events.

Figure 7 Annual maxima events (Q_max) and their 30 d pre-event effective rainfall at Maastricht (a) and Vaals (b), including their preceding extreme indicators. The top five floods during the study period are shown with their year of occurrence. The dashed purple line represents the linear fit, using the least-squares approach, between the effective rainfall of the high-flow events preceded by P₉₉ and P_MD and their respective Q_max values, while the red dashed line represents the linear fit between the effective rainfall of the Compound-III events and their Q_max. The shaded areas show the 95 % confidence intervals for the fits, and the Pearson correlation coefficients (ρ) are also reported.

Figure 8 Multi-temporal trend analysis for P_1D, P_5D, P_15D and P₉₅ at Vaals for the (a) winter half-year and (b) summer half-year. Each pixel presents a fixed single period (minimum window length of 30 years) of the start and end year. For each period, the M–K test is applied, and the color indicates the Z-statistic value of the test (the same definitions apply to the subsequent figures). Blue colors indicate increasing trends, whereas red colors indicate decreasing trends; the darker the color, the more significant the trend. Statistically significant trends are considered to be those with Z-statistic values higher than 1.28 (or smaller than −1.28 for downward trends) corresponding to the defined significance level of 0.2 (see Sect. 2.4).

3.2 Trend analysis

3.2.1 Flood driver trends

Figure 8 illustrates the multi-temporal trend analysis for several precipitation indices for the Vaals station, for half-year periods, as an example. The multi-temporal analysis for Vaals, which has a record from 1952 to 2021, results in 861 trends. In the winter half-year, statistically significant increasing trends are found for the longest periods in all indices at Vaals. However, a decreasing, mainly insignificant, pattern is visible in the recent past (trends starting after the 1980s). In summer (Fig. 8b), negative trends are visible for the longest periods, while this changes to positive trends in the recent past for k ≤ 5 d. Summer trends for k > 5 d are rather mixed: generally insignificant trends are found, with shifts between positive and negative tendencies. For the full multi-temporal analysis per index and station, the reader is referred to the Supplement.

Figure 9 shows the consistency in statistically significant trends in each precipitation index per rainfall station. In winter, only increasing trends are visible, with the exception of the P_15D index at Ubachsberg. The decreasing tendency in the recent past detected at Vaals is very strong and statistically significant at Ubachsberg for most indices (Fig. S3), causing this inconsistent decrease. Consistent and strongly consistent increases are observed for at least one station per index. For the P_7D index, four out of five stations show consistent or strongly consistent increases. For k ≤ 10 d, most stations have strongly consistent or consistent upward directions. With increasing k (15,…,40 d), the increase becomes weaker (inconsistent) for the majority of the stations; however, two out of five stations (located inside the catchment) still show increases. Indices for k ≤ 10 d show strongly consistent increases, while a consistent increase is visible for k > 10 d at the Valkenburg station. In addition, three out of five stations show strongly consistent or consistent increasing trends in the P₉₅ and P_MD indices. Overall, the trend analysis in winter for the Geul catchment shows a consistent increase in very wet days (P₉₅) and maximum k-day precipitation sums. The rise in severe precipitation is caused mostly by more rain on already wet days. Multiday precipitation extremes are consistently increasing. This is a crucial finding, as the effects of prolonged, heavy storms in combination with wet antecedent conditions appear to be the dominant flood drivers in the Geul catchment.

Figure 9 Temporal consistency of precipitation trend indices for the winter and summer half-years. Blue colors indicate upward trends, whereas red colors indicate downward trends.

Trends in summer periods show variability across the different stations. The majority of the summer half-year k-day and P₉₅ indices are subject to inconsistent trends. Most stations show generally insignificant trends, with changes between positive and negative tendencies. The only consistent trends are mainly increasing and are found at the Valkenburg station for k = 5, 7, 10 and 15 d; at Ubachsberg for k = 7 d; and at Maastricht (i.e., P_MD ). In addition, the P_30D and P₉₅ indices at Maastricht show strong and statistically significant increasing trends in the majority of tested cases, resulting in weakly consistent trends. As mentioned, despite the fact that the difference between significant increasing and decreasing trends in the summer half-year is not clear, the statistically significant increasing trends in the recent past, mainly for k ≤ 3 d, are strong and should be taken into account (Figs. S6–S10). In addition, the consistent increasing trends at Valkenburg reveal a direction towards more wet conditions in the summer half-year.

Figure 10 Half-year yearly maxima time series (a) and multi-temporal trend analysis for the (b) winter half-year yearly maxima (Q_W,max) and the (c) summer half-year yearly maxima (Q_S,max). The gray stripes in the heatmaps indicate the excluded hydrological years (i.e., 1971, 1974 and 1990).

3.2.2 Discharge trends

The results of the multi-temporal analysis for the Q_W,max and Q_S,max time series are shown in Fig. 10. It can be observed that the maximum flows show variability over the two half-year periods. Increasing trends are found in the longest periods for the winter half-year, but this seems to have changed in the recent past to statistically insignificant decreasing tendencies. Overall, the increase in the Q_W,max is considered to be consistent when accounting for the missing hydrological years of 1971, 1974 and 1990. This pattern is in agreement with the extreme precipitation trends in the area for winter, as large similarities are observed in terms of magnitude, direction and variability. For example, Q_W,max trend variability is quite similar to the P₉₅ index in winter at Vaals (see Fig. 8a): statistically stronger increasing trends are found over longer periods (from 1970 to 2021), with a weak decreasing direction in the recent past (trends starting from the 1980s).

Mixed and statistically nonsignificant trends are observed in the summer half-year (Fig. 10c), as expected, considering that the trends in extreme precipitation in the same period are inconsistent and their strength is (statistically) insignificant. For some cases, Q_S,max trends shift between negative and positive tendencies in similar periods with extreme summer precipitation; however, this match is not as clear as in the winter period. The increasing direction of extreme precipitation in the recent past for summer starts also becoming visible in the Q_S,max direction (see positive values for trends starting after 1985). In general, the effect of the considerable increase in evaporation in the area in summer (Tsiokanos, 2022) and the large soil moisture deficits should be taken into account when translating extreme precipitation to extreme streamflows, and subsequently discussing correspondences and differences between discharge and precipitation trends.

4 Discussion

4.1 Data uncertainty

Records of 24 h precipitation from the KNMI manual rain gauge network are employed. The used precipitation time resolution may be considered low for flood analysis; however, the response time (i.e., longer than a day; see Sect. 2.1) of the catchment allows for the application of this resolution. In addition, the main goal of this paper is to investigate the role of P₉₉, P_MD and P_WAC as potential flood drivers. The use of a 24 h resolution can affect the defined P₉₉, whereas the applied resolution does not have a major impact on the definitions and meaning of P_MD and P_WAC. For this reason, allowances were made with respect to the way that we defined a P₉₉ day (i.e., precipitation amount on the same or previous day with respect to the high-flow event; see Sect. 2.3.3).

Long precipitation time series may have been influenced by instrumental modifications and station relocations throughout the recording period. As mentioned in Sect. 2.2, the data are considered to be of high quality, as KNMI performs regular quality tests. In addition, two homogeneity tests were applied to monthly sums. In general, it is assumed that the analyzed precipitation time series in this research are not affected by instrument or location alterations, so the trends that are found can be attributed to climate and not to human interventions.

It must be acknowledged that the produced discharge results are subject to significant uncertainty (Di Baldassarre and Montanari, 2009). Estimations during extremely high flows are very inaccurate. For example, the recorded discharge during the July 2021 flood was 55 m³ s⁻¹, whereas it is estimated that it actually exceeded 80 m³ s⁻¹ (van Heeringen et al., 2022). However, our main findings regarding the role of compound events in generating high flows remain valid. In addition, periods of transition; changes in gauge position, equipment and monitoring frequencies; and stage–discharge relations can cause sudden variations in flow rates. These changes can be more visible in mean flow trends, the values of which are low in contrast to the high flows used in this paper. The long-term measurements at the Meerssen station at the outlet of the catchment are considered reliable in terms of homogeneity (Agor, 2003). The high-flow trends are found to be similar to the direction and significance of the extreme precipitation trends, indicating that the results are likely not affected.

4.2 Implications

For the Geul, it is found that extreme daily precipitation alone is typically insufficient to cause floods. However, wet antecedent soil conditions are a crucial factor determining the probability of flooding. In this respect, the finding that heavy, prolonged precipitation frequently preceded high flows in the Geul seems reasonable, as multiday precipitation can also serve as a proxy for heavy precipitation occurring under wet antecedent circumstances (Nanditha and Mishra, 2022). Most of the flood events are observed in winter periods, when the catchment tends to be very wet, with shallow groundwater tables. In summer periods, most of the extreme (intense) precipitation events are not translated to high-flow peaks, due to large soil moisture deficits. The most devastating flood in the area (i.e., July 2021) was aggravated by rainfall events in the preceding days and weeks. The role of wet antecedent (soil) conditions in driving floods is well established; however, to date, the focus has tended to be on larger catchments (Wasko et al., 2020). Information about the initial (wet) conditions of the catchment is deemed essential, particularly for flood forecasting, as the local water authority does not currently monitor soil moisture. In addition, the geology of the Geul region can significantly control the runoff response, as there is a thick unsaturated chalk zone that can store much water (Klein, 2022). Therefore, the (geo)hydrological properties and characteristics of the catchments should receive more attention in the flood forecasting system (Zanon et al., 2010; Douinot et al., 2022). Our findings are also expected to help with understanding flood mechanisms in other lowland or chalk catchments around the world. In addition, the event-based approach followed in this work can be exploited in other catchments to examine the relative role of wet antecedent soil moisture conditions and precipitation characteristics preceding high flows, especially in areas where (long-term) soil moisture data are not available.

The statistical results obtained in Sect. 3.2.1 demonstrate some intriguing variations in the Geul catchment's precipitation regimes across the studied periods. The most notable change is the consistent and strong increase in critical precipitation during the winter half-year. During this period, various indices representing heavy, prolonged events, such as P_kD for k ≥ 3 d and/or P_MD, as well as 24 h extreme indices, like P_1D and P₉₅, show mostly consistent increases. These combinations of indicators can contribute to the saturation of the catchment, thereby increasing the risk of flooding. In addition, it appears that a portion of the rise in severe precipitation stems from increased rainfall on already wet days, as evidenced by consistent (or strongly consistent) rises in P_3D, P_5D, P_7D and P_MD across the majority of stations. All of these findings are crucial, as heavy and prolonged storms in combination with wet antecedent conditions have impacted the catchment and caused floods, mainly in winter. Although it cannot be concluded that climate change had a significant impact on the July 2021 flood event in the Geul region, as there are no apparent consistent patterns in most summer precipitation trends, a concerning increasing direction in the recent past (mainly after the 1980s) is visible. This finding is important, as it shows that (except for the intense showers in summer) the effects of heavy storms in combination with wet antecedent conditions should be taken into account. At the same time, it is critical to consider the substantial increase in summer potential evaporation rates (due to increases in temperature and radiation) in the area (Tsiokanos, 2022) that may lead to soil moisture deficits, when translating extreme precipitation events into potential extreme flows. According to the recently published KNMI scenarios, wetter winters and increased extreme summer showers are projected for the Netherlands (KNMI, 2023). These projections suggest that the number of heavy showers with significant precipitation is expected to rise, with a shift from light to heavier (more rain falls from the shower) and more intense (more rain falls in a certain period of time) showers (KNMI, 2023). These climate scenarios are in line with the trends found in this study. Overall, the long-term trends in the critical precipitation are also visible in the runoff patterns. Thus, the effect of climate change should be seriously considered in the area and should be incorporated into flood designs; moreover the effects of agro-economic developments, such as land-use changes, should also be taken into account. Furthermore, our findings from the precipitation and discharge trend analyses can serve as a valuable reference for assessing the impact of climate change on precipitation and discharge patterns in regions other than the Geul. In addition, the use of a multi-temporal approach, including the consistency criteria, appears useful and is recommended for identifying variability, recent directions and long-term trends.

5 Conclusions

We used an event-based approach to detect the main flood drivers in the Geul River catchment and a multi-temporal trend analysis to investigate their temporal variability and consistency. Our results indicate that heavy multiday precipitation can have a notable impact on high flows, preceding them approximately 75 % of the time (using a 4 d precipitation duration; Fig. 4). Nevertheless, wet antecedent conditions play a crucial role in translating extreme precipitation events into extreme flows and make the difference between a “regular high flow” and a flood event. Extreme 24 h precipitation without wet antecedent conditions, which mainly occurs in summer, has never independently led to floods in the past (see Fig. 5 and Table 3). The joint probability of extreme (prolonged) rainfall and wet initial conditions (which can be seen as a compound event) determines the chances of flooding. As a result, prolonged, heavy rainfall and wet antecedent wetness appear to be the primary factors contributing to extreme discharge events, and they should be used as flood indicators, rather than extreme precipitation alone. Flood-producing precipitation shows a consistent increase in the winter half-year, a period in which more than 70 % of extremely high flows have historically occurred. Heavy, prolonged storms in combination with wet antecedent conditions can cause large floods, and these conditions are becoming more frequent during winter. This rise is also reflected in the winter half-year maximum discharges, which are increasing in terms of magnitude. Although the majority of precipitation and flow trends do not exhibit consistent patterns in the summer half-year, a notable and concerning upward direction has become evident in the recent past. This observation underscores the necessity to account for the effect of compound events in addition to intense summer showers. The extreme flood event in July 2021, along with the observed increase in flood drivers, emphasizes the importance of incorporating compound events into flood risk assessment.