The threshold-level approach has traditionally been applied to detect hydrological drought events (Van Loon, 2015; Heudorfer and Stahl, 2017; Brunner and Chartier-Rescan, 2024), where an event is defined when streamflow falls below a given threshold. To account for seasonal patterns, the use of variable thresholds, such as seasonal, monthly or daily, has been preferred over using a unique fixed threshold (Brunner et al., 2022; Van Loon and Laaha, 2015). On the other hand, floods are typically detected in a time series by using either a block maxima method or the partial duration series sampling method, which captures values above a threshold (Pan et al., 2022). Although less common, the variable threshold-level approach used for drought identification can also be effectively applied to characterize HFS in a consistent and unified manner (Quesada-Montano et al., 2018). However, in practice, when automatically identifying HFS, its application can result in identifying days above the threshold during the low flow season that actually do not represent high flow conditions in the catchment. Therefore, in this study, we adopt a mixed threshold approach (Caillouet et al., 2017). We select as threshold for a given day the minimum (for LFS) or maximum (for HFS) value between a quantile-based fixed threshold and the monthly threshold associated to that day. This mixed threshold approach is more restrictive for spell detection, while still accounting for some seasonal patterns.

The inclusion of baseflow is another characteristic of the method adopted. When analyzing hydrologic floods or drought events, independence between events is often ensured by setting a minimum number of days between occurrences (Diederen et al., 2019; Brunner et al., 2020, 2021; RahimiMovaghar et al., 2024), by using pooling procedures (Hisdal et al., 2024) or by smoothing the time series with a moving average filter (Fleig et al., 2006). In our study, the focus is on ensuring independence between consecutive spells. While the threshold exceedance can identify the starting date of a spell, we introduce the baseflow as a criterion to determine its end date. The idea is that baseflow can be used as a proxy for determining when the catchment returns to normal conditions, effectively pooling events that may form a spell.

We estimated baseflow with three different methods: the UKIH smoothed minimum method (UKIH, 1980), the Pelletier-Andréassian (PA) conceptual model (Pelletier and Andréassian, 2020), and the Lyne-Hollick (LH) recursive digital filter (Lyne and Hollick, 1979). The UKIH method identifies turning points from minimum daily streamflow values within non-overlapping blocks (five days by default), and estimates the baseflow through linear interpolation (Piggott et al., 2005; Laaha and Koffler, 2022; Hisdal et al., 2024). The PA method represents the baseflow as the delayed outflow of a conceptual quadratic reservoir (Pelletier and Andréassian, 2020; Pelletier et al., 2020). The LH method is based on signal decomposition analysis with one filtering parameter to separate the low frequency baseflow component from the high frequency quickflow component (Duncan, 2019). We applied it following the guidelines proposed by Ladson et al. (2013), using the recommended parameters for daily data: a filter parameter of 0.925 with 3 passes (forward-backward-forward). Default parameter settings were also used for UKIH and PA methods.

Figure 1 illustrates the spell detection procedure applied to HFS (panel a) and LFS (panel b). First, the mixed thresholds are calculated: the figure illustrates the threshold based on the exceedance of the 99th percentile (q99) for HFS (blue dotted line) and the threshold based on the non-exceedance of the 20th percentile (q20) for LFS (red dotted line). The start date (ti) of a HFS (LFS) spell is identified when the streamflow is above (below) the mixed threshold (Qt). Once the baseflow is calculated (green dashed lines), it helps to define the spell end date. For HFS, we first evaluate the date when the streamflow falls below the threshold and equals (or, depending on the baseflow method used, is close to) the baseflow (te,b). We then search backwards the last date when the streamflow remained above the threshold Qt, which marks the end of the spell (te,f). For LFS, the date te,b is defined when the baseflow exceeds the threshold Qt. The spell end date (te,f) is identified by checking backwards the last date when the streamflow was below the threshold Qt. To account for instances where streamflow temporarily exceeds the threshold for a few days in low flow condition, an additional step is introduced for LFS detection, which pools together spells if there is only five days or less between them. Finally, the automatic procedure verifies that there is no HFS within the LFS.

We characterize spells with catchment-specific thresholds using two sets of percentiles for the mixed threshold: one more severe, based on the 99th and the 5th percentiles for HFS and LFS, respectively, and one less severe, based on the 95th and the 20th percentiles for HFS and LFS, respectively. To avoid very short LFS that would not be associated with impacts we excluded the LFS with total duration smaller than seven days. For each spell, we calculated the spell duration (difference between start and end dates), the cumulative water deficit (surplus) for LFS (HFS) (in mm) (a measure of severity), the percentage of time the HFS (LFS) spell remained above (below) the threshold, and the number of times the threshold was crossed during the spell duration (which allows us to estimate the number of pooled events within the spell).

For a more robust analysis, we grouped the spells based on their duration, defining the spell categories shown in Table 1. HFS that last up to three days are classified as short duration (HFS_S), between three and 15 d as medium duration (HFS_M), and over 15 d as long duration (HFS_L). For LFS, short duration spells (LFS_S) last between seven and 30 d, medium duration spells (LFS_M) last between 30 and 90 d, and long duration spells (LFS_L) exceed 90 d. Hereafter, we use a generalized notation where *_S, *_M, and *_L refer to short, medium, and long duration spells, respectively, with the asterisk (*) representing either HFS or LFS.

Four types of spell transitions can be assessed: (i) transition from a low flow spell to a high flow spell (LFS-HFS), (ii) transition from a high flow spell to a low flow spell (HFS-LFS), (iii) transition between two consecutive low flow spells (LFS-LFS), and (iv) transition between two consecutive high flow spells (HFS-HFS). Within each transition type, different spell duration categories (short, medium, long) can also be considered. In our study, only transitions with continuous data are analyzed (i.e., if there are missing data between spells, their transition is not considered in the analysis). Figure 2 illustrates the approach: three short duration high flow spells (HFS_S), one medium duration low flow spell (LFS_M), and one medium duration high flow spell (HFS_M) are detected; three transitions are defined (yellow blocks) for the time span between the first two HFS_S (t1), the third HFS_S and the LFS_M (t2), and the LFS_M and the HFS_M (t3); the transition between the second and the third HFS_S is not considered since streamflow data is missing.

The duration of a transition is quantified in days, measured from the end date of the first spell to the start of the second spell.

Our analysis focuses on two categories of interest: transitions with duration less than or equal to 30 d (hereafter, within-a-month) and those with duration between 31 and 90 d (hereafter, seasonal). The remaining transitions, those exceeding 90 d, are also quantified but fall outside the main focus of our study, which emphasizes timescales relevant for short-to-medium term of the disaster risk management cycle, where impacts of consecutive events can be amplified before recovery from an earlier event is complete (De Ruiter et al., 2020; Ward et al., 2026). The within-a-month category also includes the rapid transitions (less than 14 d) as defined by Götte and Brunner (2024).

To quantitatively compare the occurrence of transitions across flood forecasting centers (hereafter referred to SPC for “Service de Prévision des Crues” in French), we analyze the frequency of within-a-month and seasonal transitions. For each SPC, we calculate the transition frequency as an exposure-weighted annual rate (i.e., ratio of total transitions and total years of data), so that catchments with longer streamflow data availability contribute proportionally to the regional frequency estimate.

To assess whether a specific region of interest experiences significantly more or fewer transitions than the national average, we calculate the standard score (z-Score) for each SPC relative to the distribution of frequencies across all SPCs. To account for potential uncertainty in the calculation of the standard score due to the relatively low frequency of these short-to-medium term transitions, we conduct a bootstrap procedure with replacement (10 000 iterations). In each iteration, catchments are resampled, and the standard score is recalculated for each bootstrap sample to generate an empirical distribution of z-scores. We estimate the 95 % confidence interval for each SPC from the bootstrap distribution using the bias-corrected and accelerated bootstrap method, which adjusts for bias and skewness in the bootstrap distribution (Davison and Hinkley, 1997).

To analyze when the transitions mainly occur within the year (i.e. transition seasonality), we evaluate, for each transition, the date corresponding to its midpoint, which is calculated by adding half of the transition total duration to the end date of the first spell. We utilize circular statistics (Ley and Verdebout, 2017) to analyze the transition timing, adapting methodologies established in flood seasonality research (Hall and Blöschl, 2018; Tramblay et al., 2023; Bagheri-Gavkosh and Hosseini, 2023; Fang et al., 2024). Transition dates are converted to angular values to calculate the mean transition date and the concentration index. This index measures the variability of the transition dates around their mean value. To ensure meaningful interpretation of the transition seasonality, we applied two circular uniformity tests with complementary properties. The Rayleigh test (Mardia and Jupp, 2000) detects unimodal departures from uniformity by testing whether the mean resultant vector length is significantly greater than zero (Ley and Verdebout, 2017). The Hermans-Rasson test (Landler et al., 2019) is sensitive to any departure from uniformity including multimodal distributions (García-Portugués and Verdebout, 2018). Both tests are implemented through the sphunif R package (García-Portugués et al., 2024). Only stations where both tests reject the null hypothesis of circular uniformity (p<0.05) are included in the analysis, ensuring that mean transition dates correspond only to unimodal distributions.

We apply our methodology to streamflow data from a set of 643 catchments from the CAMELS-FR dataset in France (Delaigue et al., 2024, 2025), covering a large range of hydroclimatic conditions (from oceanic to continental, mountainous and Mediterranean conditions) (Fig. 3). Catchment mean elevation ranges from a lower quartile of 173 meters above sea level (m a.s.l.), over a median of 336 m a.s.l., an upper quartile of 695 m a.s.l. to a maximum of 2703 m a.s.l., and catchment areas range from a lower quartile of 98 km² over a median of 189 km², an upper quartile of 473 km² to a maximum of 110 188 km². Since the threshold level approach is not suitable for drought identification in intermittent rivers (Van Loon, 2015; Sarremejane et al., 2022), we only considered CAMELS-FR catchments that had less than 5 % of days with zero flow. In addition, streamflow data flagged as questionable by the producer for 30 consecutive days or more within the less severe threshold zone, i.e. when streamflow is above the 95th percentile and below the 20th percentile, were removed from the time series. Overall, daily streamflow data are available over the period 1970–2021, with each catchment having at least 25 complete hydrological years, defined as a year with less than 20 % missing data. The number of complete hydrological years per catchment ranges from a lower quartile of 37 years, over a median of 45 years, and an upper quartile of 50 years.

To aggregate the results geographically, and illustrate the potential of the temporally compound event analysis to operational hydrology, we rely on the SPCs operated by the French national service (SC Vigicrues), comprising 17 centers at the time of writing, as the spatial units for our regional analysis. For geographical context, these SPCs are presented within six major hydro-administrative regions for hydrological purposes (Fig. 3).

The LH and UKIH methods had similar results in the detection of spells, with Pearson correlation coefficients of 0.99 for HFS and 0.89 to 0.92 for LFS (Fig. A1; see Appendix A for the pairwise comparison). The main difference between them is in how the baseflow end-date is identified. For HFS, UKIH consistently identified it earlier than LH, resulting in slightly more frequent but shorter spells. For LFS, this behavior was mixed across catchments. Despite these differences, both methods produced consistent distributions of spell duration (Fig. 4a) and of the percentage of time streamflow remained above (HFS) or below (LFS) the threshold (Fig. 4b).

The PA method showed a distinct behavior: despite implementing a 10 % tolerance criterion between streamflow and baseflow, PA's baseflow estimates rarely converged with the observed streamflow for HFS, delaying the identification of the baseflow end-date (te,b) and merging multiple HFS into single prolonged spells (Fig. 4a). PA therefore correlated less with LH and UKIH for HFS, especially under the less severe threshold (r= 0.72 and r= 0.71), and gave lower percentages of time above the threshold for medium and long HFS (Fig. 4b). For LFS, PA correlated more closely with LH and UKIH under the more severe threshold (r= 0.79 and r= 0.82, respectively), but still overestimated LFS duration under the less severe threshold (Fig. 4a).

Figure 5 illustrates the methodological differences between the LH and UKIH methods in two case studies. The first case is the major 2001 floods in the Somme River basin in northern France, located in the Bassin du Nord hydro-administrative region (Fig. 3a) on the Nièvre River (affluent of the Somme River) (Habets et al., 2010). The second case is the 1996 hydrologic-agricultural drought that hit north and west of France (Barraqué et al., 2010), represented by the Yères River basin, located in the Seine-Normandie hydro-administrative region (Fig. 3a), in the SPC Seine aval-Côtiers Normands (Fig. 3b). In the first case, the LH method detected one continuous long duration HFS, while the UKIH method identified 10 separate shorter spells. In the second case, the LH method detected one continuous long duration LFS in 1996, while the UKIH method split this into one LFS_M and one LFS_L, with both methods following with a LFS_L in 1997. In both cases, a single prolonged spell better reflects the actual hydrological conditions than multiple shorter events. For extended HFS in particular, the UKIH method's fragmentation can misrepresent prolonged flood conditions in groundwater dominated catchments. Given the overall good agreement between LH and UKIH methods across spells metrics (Fig. 4), and the better ability of LH method to detect long duration spells, the LH recursive digital filter method is used for the subsequent analyses in this study.

The use of catchment-specific percentiles allows for the detection of HFS and LFS in each catchment where observed time series of streamflow are available, accounting for regional hydrological variability. Through the application of this methodology to our dataset of 643 catchments in France, we detected 145 721 spells in total across all catchments, with 71.9 % classified as HFS and 28.1 % as LFS. As expected, the majority of spells fell under the less severe threshold category. Within this category, 77 347 HFS were identified, with 54.8 % being of short duration, 35.0 % of medium duration, and 10.2 % of long duration. Similarly, 31 849 LFS were observed in the less severe threshold category, comprising 63.8 % of short duration, 24.3 % of medium duration, and 11.9 % of long duration. On average, each catchment experienced approximately 2.7 HFS yr⁻¹ (first quartile: 2.1 HFS, third quartile: 3.2 HFS), and 1.1 LFS yr⁻¹ (first quartile: 0.95 LFS, third quartile: 1.3 LFS).

For the more severe threshold category, 36 525 spells were identified, comprising 27 392 HFS and 9133 LFS. Within this category, short duration HFS were more common, accounting for 79.2 % of all HFS, while long duration HFS constituted only 3.4 %. Among LFS, 76.1 % were in the short duration category, 19.8 % were of medium duration, and 4.1 % of long duration. This corresponds to an average of approximately 0.96 HFS per catchment per year (first quartile: 0.68 HFS, third quartile: 1.2 HFS), and 0.32 LFS per catchment per year (first quartile: 0.26 LFS, third quartile: 0.37 LFS). Figure 6 illustrates the spatial distribution of HFS and LFS frequencies across France by spell severity (less severe, more severe), and duration (short, medium, long).

The more severe the threshold and the longer the duration, the lower the frequency of spells per year. For the most restrictive case (more severe and long duration), our methodology results in 46 % and 73.4 % of the catchments with at least one spell for LFS_L and HFS_L, respectively. For all the other cases, there are 95 % or more of the catchments with at least one spell. When considering catchments with a minimum of occurrences of five spells, the spatial coverage significantly decreases, especially for the more severe thresholds and long duration spells, with only 3.4 % of catchments represented for HFS and no catchment for LFS. This indicates the rarity of extreme and prolonged spells in our case study (catchment dataset and study period).

Interestingly, only two catchments lacked short duration HFS, both located in the Bassin du Nord hydro-administrative region (Fig. 3). This region is characterized by slow response catchments sustained by chalk aquifers (Habets et al., 2010). The more severe HFS of longest duration recorded by our methodology, with a total duration of 112 d, was in this region (2001 flood in the Somme River basin, shown in Fig. 5a). Between autumn 2000 and spring 2001, northwestern France experienced exceptional pluvial conditions, leading to flooding of the Somme River and its tributaries, as well as groundwater flooding, where underground water overflowed onto the plateau, particularly in the chalky areas (CCR, 2021). The 2001 flood event in the Somme River basin was notable for its prolonged duration and the extended time required for the hydrologic system to drain, leading to elevated water levels for at least three months in most areas and up to six months in some regions of the catchment (Pennequin, 2010). In our methodology, multiple individual events occurring in close succession can be pooled into a single spell based on the catchment baseflow condition. The 112 d HFS in the Somme catchment pooled 11 events (Fig. 5a). This was the maximum number of pooled events that resulted from the application of our methodology for HFS. Overall, for HFS across severity and duration categories, the median number of pooled events varied between one to four pooled events.

Figure 7 shows the mean cumulative water surplus (deficit) for HFS (LFS), calculated by summing up the daily discharge (as runoff in mm) that occurred above (below) the threshold during a spell's duration. It is important to recall that, in our spell detection methodology, a spell's total duration does not necessarily mean that the streamflow remains continuously above (for HFS) or below (for LFS) the threshold throughout its entire duration (Fig. 4b). While an increase in spell duration generally corresponds to a higher cumulative surplus (deficit), we observed for HFS that the cumulative water surplus tends to reach a plateau between HFS_M and HFS_L when using the more severe threshold (HFS_M mean: 14.6 mm, first quartile: 3.4 mm, third quartile: 17.1 mm; HFS_L mean: 12.6 mm, first quartile: 3.6 mm, third quartile: 15.4 mm). This counter intuitive result is explained by the proportion of time the spell remains above the threshold for HFS. HFS_M often remains above the threshold for a higher percentage of time compared to HFS_L (Fig. 4b), which may experience more frequent drops below the threshold. Consequently, the cumulative surplus volume does not always increase proportionally with longer durations. On the other hand, the cumulative water deficit, for LFS, reveals a more consistent increase as the duration category extends from short to long duration.

As shown in Fig. 4b, long duration LFS remains below the threshold for a high percentage of time, indicating that our methodology effectively captures persistent and sustained periods of low flows. Pooled events were more often observed in LFS detection than in HFS detection, mainly due to typical fluctuations in low streamflow near the threshold. In the more severe threshold and long duration LFS, the median number of pooled events was seven (first quartile: four events, third quartile: nine events), with a maximum of 33 events in the Vonne catchment located in the SPC Vienne-Charente-Atlantique (Fig. 3b) between 19 May 1976 to 2 November 1976. Medium duration LFS had a median of three pooled events (first quartile: two events, third quartile: five events), reaching up to 14 events. In contrast, short duration LFS mostly involved pooling only one or two events, though in one occasion up to eight events were pooled in a station located in the SPC Garonne-Tarn-Lot (Fig. 3b) between 15 November 1985 and 8 December 1985.

Although most identified LFS fell in the short duration category, our methodology effectively captures medium and long duration spells, including those exceeding six months in the more severe threshold and some exceeding one year when using the less severe threshold for spell detection. An illustrative example is the detection of the longest LFS with the more severe threshold in the Seine-Normandie hydro-administrative region (Fig. 3a), shown in Fig. 5b. It lasted 316 d, between 1996 and 1997, and was followed by another LFS_L with 182 d duration in less than four months in the same catchment. Although this LFS is not related to the most remarkable droughts in 1976 and 2003 in France, it captures the 1996 drought event. This event was characterized by anticyclonic conditions that led to exceptionally low rainfall in winter and spring in the northern regions of France, which propagated to hydrological deficits (CCR, 2018a; Parry et al., 2012), and further depleted soil storages, propagating into agricultural drought in the northern and western regions (Amigues et al., 2006; Barraqué et al., 2010). Even though heavy rainfall during the end of the year partially restored the water levels, deficits persisted into 1997, leading to the reestablishment of drought conditions with precipitation deficits ranging from 25 % to over 50 % in France (CCR, 2018b). This prolonged LFS detected in our methodology remained below the more severe threshold for 86.7 % of its duration, pooling 12 events together.

We analyzed the frequency and duration of spell transitions (temporally compound events). Transitions can occur between the same or different types of spells (Fig. 2), i.e., from HFS to HFS and from LFS to LFS, but also from HFS to LFS and from LFS to HFS. In addition, one may also consider the different thresholds (more and less severe) and durations (short, medium and long).

Figure 8 shows transition patterns between hydrological spells, allowing us to visualize the proportion of occurrence of each type of transition for three selected configurations. Firstly, we consider only HFS and LFS detected by using the less severe threshold (Fig. 8a). Secondly, only HFS and LFS detected by using the more severe threshold (Fig. 8b). Thirdly, only HFS detected by using the more severe threshold and LFS detected by using the less severe threshold (Fig. 8c).

We can see that consecutive HFS-HFS transitions are the most frequent transition type, accounting for 56 % of transitions for the less severe threshold (Fig. 8a) and 63 % for the more severe threshold (Fig. 8b). Transitions from HFS to LFS, LFS to HFS, and consecutive LFS were more equally represented, each comprising approximately 15 % in the less severe threshold and 12 % in the more severe threshold. As mentioned previously, the transition analysis has multiple dimensions. When selecting spells that were detected following a particular threshold of interest, one can obtain different patterns. For instance, the evaluation of transitions using the less severe threshold for LFS threshold and the more severe for HFS (Fig. 8c), the most frequent type of transition is given by consecutive LFS (35 % of transitions), followed by consecutive HFS (28 %). One can also be interested in examining spell transitions within only some predefined spell duration categories (short, medium and long). Figure 8 shows, for instance, a predominance of transitions between short duration spells (HFS_S and LFS_S). It also shows combinations involving short-to-medium (*_S to *_M), and medium-to-short (*_M to *_S) spells.

While Fig. 8 allows us to focus on the proportion of a particular transition configuration, such as using the less severe threshold for LFS and the more severe threshold for HFS, for instance, it does not inform on the duration of the transitions. Transition durations can vary widely, ranging from as short as 0 d to as long as 5197 d (approximately 14 years) for the more severe threshold, and up to 2176 d (about six years) for the less severe threshold. The longest transition durations for both thresholds were recorded in the SPC Bassin du Nord between two LFS for the more severe threshold, and between an HFS and LFS for the less severe threshold. Such long duration transitions can be seen as a catchment with normal flows, or with a long period without experiencing too high or too low flows that could be characterized as HFS or LFS by our methodology. Very short transitions can be seen as an indication of catchments with highly variable flows, going towards extremes in short periods of time. The shortest observed transition duration (0 d) occurred exclusively from LFS to HFS, where a low flow condition abruptly changed to a HFS. In order to investigate this further, we considered the aggregation level offered by the flood forecasting centers in France, as presented in Sect. 2.4. Figure 9 shows the proportions of spell transition duration categories for each SPC, using the less severe threshold for LFS and the more severe threshold for HFS. Transitions are categorized into within-a-month (≤ 30 d, shown in orange) and seasonal (31 to 90 d, yellow), with remaining transitions (> 90 d), in blue. The first two columns in Fig. 9, representing transitions between consecutive spells of the same type (HFS to HFS and LFS to LFS), show relatively homogeneous patterns across the SPCs. The proportion of within-a-month transitions is generally between 17 % to 25 % of consecutive HFS, and 39 % to 47 % of consecutive LFS across SPCs. On average, consecutive HFS transitions last 147 d (first quartile: 34 d, third quartile: 219 d), while consecutive LFS transitions last 110 d (first quartile: 16 d, third quartile: 183 d).

Transitions from HFS to LFS (third column in Fig. 9) are relatively slow, with a high proportion of transitions lasting more than 90 d (generally between 77 % to 91 %). On average, HFS to LFS transitions last 212 d (first quartile: 107 d, third quartile: 242 d), suggesting a gradual development of LFS after an HFS. Within-a-month transitions from HFS to LFS are rare, accounting for only 1.8 % of all HFS to LFS transitions, and occurring in only 16 % of catchments. The highest proportion of within-a-month transitions from HFS to LFS (8 %) is found in the SPC Seine moyenne-Yonne-Loing, which covers part of the Seine River basin. The SPC Rhône amont-Saône and SPC Alpes du Nord, in the upstream part of the Rhône-Méditerranée hydro-administrative region, were the only regions with at least 25 % of HFS to LFS transitions occurring within less than 90 d (the combined within-a-month and seasonal categories).

On the other hand, transitions from LFS to HFS (fourth column in Fig. 9) occur in a shorter time window. The proportion of within-a-month transitions is generally between 10 % and 32 %, while seasonal transitions are between 27 % and 36 %. On average, LFS to HFS transitions last 142 d (first quartile: 47 d, third quartile: 157 d). As observed in Fig. 9, there is a more pronounced spatial pattern for this transition type with within-a-month and seasonal transitions predominantly occurring in the Rhône-Méditerranée hydro-administrative region (e.g., SPC Alpes du Nord: 34 % within-a-month, 26 % seasonal; SPC Méditerranée Est: 37 % and 35 %, respectively). These shorter transitions are rare in the Loire-Bretagne hydro-administrative region (e.g., only 5 % within-a-month occurrence in SPC Vilaine-Côtiers Bretons and 6 % in SPC Maine-Loire aval).

To better understand the spatial patterns of these short-to-medium-term transitions, Fig. 10 shows the standardized transition frequencies aggregated at the SPC level, using the less severe LFS and more severe HFS threshold. Confidence intervals for these standard scores were obtained from bootstrapping experiments, allowing us to identify SPCs where transition frequencies are above (green) or below (purple) the national mean. The spatial patterns in Fig. 10 indicate regional contrasts. The SPC Alpes du Nord shows transition frequencies above the national mean for all transition types (z-score above two for either within-a-month or seasonal transitions). On the other hand, SPC Bassin du Nord usually shows lower frequencies in all transition types. Overall, SPCs in the Rhône-Méditerranée, Adour-Garonne, and Rhin-Meuse regions tend to have transition frequencies above the national average, while those in the Bassin du Nord, Seine-Normandie and Loire-Bretagne regions generally have lower transition frequencies.

Specific transition types reveal distinct spatial patterns. For consecutive HFS, northern and western SPCs show less frequent transitions, while SPCs in south of France show higher frequencies. For consecutive LFS, only SPCs in the Rhine-Meuse and Rhône-Méditerranée (Rhône amont-Saône, Alpes du Nord) hydro-administrative regions have transition frequency above average. Regarding HFS to LFS transitions, it is interesting to point out the SPC Seine moyenne-Yonne-Loing for ranking above average for within-a-month transitions but below average for seasonal transitions. The SPC Loire-Allier-Cher-Indre displays the inverse pattern, highlighting regional heterogeneity in transition dynamics within SPCs. For LFS to HFS transitions, SPCs in Seine-Normandie and Loire-Bretagne display transition frequencies below the national average. On the other hand, all SPCs within the Rhône-Méditerranée hydro-administrative region have within-a-month transitions frequencies above average. For seasonal transitions, SPCs in Rhin-Meuse and Adour-Garonne hydro-administrative regions, and SPCs in Rhône basin exceed the national average.

Having characterized the proportion of transition categories within each SPC and their frequency relative to the national average, we now explore the specific timescales of short-term transitions. To better analyze spatial patterns for transitions of less than 30 d (within-a-month), Fig. 11 shows maps with the median transition duration aggregated at the SPC level for the less severe LFS and more severe HFS threshold, while the proportion of within-a-month transitions between different spell duration categories is shown in the Appendix B (Fig. B1). In Fig. 11, four maps are presented, one for each transition type (panels a–d). The first panel shows transitions between HFS; the second, between LFS; the third between HFS and LFS; and the fourth between LFS and HFS. Each map is accompanied by a histogram that shows the distribution of within-a-month median transition duration for all 643 catchments.

For consecutive HFS (Fig. 11a), median values of within-a-month transition duration ranges from 15 d in the SPC Seine amont-Marne amont to 22 d in the SPC Méditerranée Ouest, and the transitions involving short duration HFS (HFS_S to HFS_S) are predominant in all SPCs (Fig. B1).

For consecutive LFS (Fig. 11b), there is less spatial variability across SPCs, with median within-a-month transition duration ranging from 14 to 16 d (the latter in the SPC Méditerranée Ouest). These transitions are mainly between short duration LFS (LFS_S to LFS_S), followed by transitions of short-to-medium (LFS_S to LFS_M), and medium-to-short (LFS_M to LFS_S), in a consistent way across all SPCs (Fig. B1).

The spatial variability of within-a-month median transition durations between HFS to LFS is shown in Fig. 11c. As identified from results in Fig. 9, these within-a-month transitions are rare (less than 9 % across SPCs). Several SPCs (e.g., Méditerranée Ouest, Bassin du Nord, Vienne-Charente-Atlantique) have fewer than four occurrences, with only one occurrence in SPC Grand Delta, and none were observed in any catchments of the SPC Seine amont-Marne amont. Despite the low transition count, some SPCs show short within-a-month median transition duration, such as 10 d in SPC Méditerranée Est. For the SPCs where these transitions occur, they are predominantly from short-duration HFS (HFS_S to LFS_*) (Fig. B1). The rapid transitions occurring between HFS to LFS are mainly preceded by a LFS in the previous two weeks (70 % of rapid transitions between HFS to LFS). For example, a tributary of the Adour River (in SPC Gironde-Adour-Dordogne), recorded in 2003 a medium duration LFS that was interrupted by a HFS, which then transitioned back to LFS in 10 d. Other examples similar to this were found across all six major administrative regions for hydrological purposes.

The spatial variability for transitions from LFS to HFS within-a-month (Fig. 11d) is characterized by shorter transition durations in the Mediterranean SPC regions, with a median transition duration as short as eight days in the SPC Méditerranée Est (Corsica included) and in the SPC Méditerranée Ouest, followed by SPC Grand Delta (12 d) and SPC Alpes du Nord (13 d). Rapid transitions, occurring within two weeks, were also observed in the Rhin-Meuse hydro-administrative region (SPC Rhin-Sarre and SPC Meuse-Moselle, with median values of 12 and 13 d, respectively). The SPCs with the highest median transition durations are concentrated in the Loire-Bretagne and Seine-Normandie regions, with median values between 17 and 22 d. In this group of transitions, we mostly encounter transitions to short duration HFS (LFS_* to HFS_S), and, to a lesser extent, to medium duration HFS (LFS_* to HFS_M). Transitions to long duration HFS (LFS_* to HFS_L) are more observed in the Seine-Normandie hydro-administrative region (Fig. B1).

Transitions can occur in different times of the year. To investigate their seasonality across space, we evaluated the midpoint date of each transition and analyzed its timing using circular statistics. Figure 12 presents the mean transition date for within-a-month transitions, using the less severe LFS and more severe HFS threshold combination. Figure 12 is divided into four panels, one for each transition type, and in each panel, the color indicates the mean transition timing (month). The mean seasonality month is only indicated for stations that reject the circular uniformity tests hypothesis (p-value < 0.05) and have at least two transitions.

When considering within-a-month spells transitions, we observe that consecutive HFS transitions (Fig. 12a) predominantly occur during winter months (December to March). Among the catchments showing significant seasonality for this transition, approximately 90 % show this winter timing, mainly concentrated in northern and western France. These transitions show moderate to high concentration indices, with an average of 0.83 (minimum: 0.46, first quartile: 0.75, third quartile: 0.94). On the other hand, consecutive LFS transitions (Fig. 12b) mainly occur in the summer (July to September) across all French territory, and also show moderate to high concentration indices, with an average of 0.74 (minimum: 0.43, first quartile: 0.67, third quartile: 0.83).

For within-a-month transitions from HFS to LFS (Fig. 12c), the seasonality is mainly statistically insignificant, with only two stations showing a significant mean seasonality, one in May (in SPC Gironde-Adour-Dordogne), and one in June (in SPC Seine moyenne-Yonne-Loing). Finally, transitions between LFS to HFS (Fig. 12d) occur primarily in the autumn and early winter (October to December), particularly in the Mediterranean and Rhin-Meuse regions. These transitions show moderate to high concentration indices, with an average of 0.86 (minimum: 0.56, first quartile: 0.8, third quartile: 0.96), indicating a well-defined seasonality pattern.

Our approach for detecting high flow spells (HFS) and low flow spells (LFS) for hydrological compound event analysis relies on daily streamflow data only. It is characterized by the integration of a mixed monthly threshold method to separate high and low flows in the time series, and a baseflow estimation method to bring together events that form unique spells.

The use of a mixed monthly threshold method allowed a more robust detection of spells. For HFS, it allowed the detection of spells that deviate from normal seasonal patterns during high flow periods, or of spells that were intense enough during low flow periods to surpass the fixed annual threshold. For LFS, it detected spells deviating from normal seasonal patterns, primarily during the low flow period, consistent with the method proposed by Caillouet et al. (2017). It must be noted that, while the monthly thresholds effectively captured seasonal variability, they can occasionally introduce artificial “staircase patterns” (Heudorfer and Stahl, 2017) when transitioning from one month to another. This can potentially affect the continuity of the detected spells, with abrupt terminations of spells, and a possible solution could be implementing a smoothed daily threshold to further enhance the representation of spell end points.

The baseflow separation method intercomparison showed that the LH and UKIH methods produce consistent results. Because spell start dates depend only on threshold exceedance (non-exceedance) for HFS (LFS), differences between methods come only from how they identify baseflow end dates. For the LH and UKIH methods, the backward search to the last threshold crossing often converges to the same spell end date. The detected spells differ only when one method identifies an earlier baseflow end date (te,b; see Fig. 1), splitting into two or more spells what the other method detects as a single continuous spell. This pattern occurred mainly in groundwater dominated catchments with long duration spells, where the UKIH method's five-day block size is likely too short to capture slow recession dynamics. In such catchments, block sizes up to 60 d may be needed to better represent baseflow response (Stoelzle et al., 2020).

Capturing the full extent of long duration spells is particularly important for groundwater flooding, which is often overlooked in traditional flood inundation modeling (Collins et al., 2020), despite its distinct impact characteristics, causing severe economic losses due to durations extending from weeks to several months (Kreibich and Thieken, 2008). By preventing spell fragmentation, our baseflow-based approach can better identify threats through a mechanism similar to preconditioned compound events (Zscheischler et al., 2020), where even a moderate rainfall during an ongoing spell can rapidly develop into a hazardous condition because the catchment has not yet recovered to normal baseflow conditions. This mechanism is supported by Berghuijs and Slater (2023), who found that antecedent baseflow conditions often have a stronger and longer lasting influence on flood magnitude than either antecedent soil moisture and short-term extreme precipitation. The importance of baseflow in determining catchment response is further reinforced by Brunner et al. (2025), who demonstrated that, across European catchments, high baseflow indices and storage properties were found to dampen the propagation of meteorological dry to wet transitions to hydrological ones, preventing rapid transitions from dry to wet states. Other baseflow estimation methods, such as the Chapman-Maxwell (Chapman and Maxwell, 1996), the Eckhardt (Eckhardt, 2005), the delayed-flow separation (Stoelzle et al., 2020), and the AutoVL (Lyne, 2025), could be explored in future applications to other datasets or regions, to further evaluate the robustness of integrating baseflow into spell detection.

The temporal resolution of the input data can also impact spell detection (Anderson et al., 2025). In small catchments with flashy hydrological response, daily mean streamflow can attenuate intense, short-duration flood peaks, placing them below the detection threshold. This can particularly impact the detection of short duration HFS and the characterization of rapid transitions involving these spells (see Sect. 4.3). It is expected, though, that the detection of LFS, whose timescales range from weeks to several months, and the characterization of medium and long duration HFS would be minimally impacted.

Finally, the selection of thresholds influences both the number of detected spells and their transitions. This threshold sensitivity was also observed by Götte and Brunner (2024), who noted that the number of drought-to-flood transitions decreases both with an increase in the flood threshold and a decrease in the drought threshold. In our study, most transitions occur between short duration spells or between short and medium duration spells, and the frequency of transitions involving long duration spells is impacted by threshold selection, resulting in small sample sizes, especially under the more severe threshold, which limits the robustness of the statistical inference for these events.

We classified spells into short, medium, and long duration categories to facilitate their characterization and analysis. As noted by Götte and Brunner (2024), there is currently no consensus on how to define or characterize transitions between hydrological extremes, and, by extension, how to categorize spell durations. Our duration thresholds of 3 and 15 d for HFS and 30 and 90 d for LFS were selected after preliminary analyses. We aimed to balance statistical robustness and hydrological relevance, providing comprehensive catchment representation in all categories. While the selection of specific duration thresholds is inherently subjective, our categorization allowed us to effectively capture the range of event durations observed across French catchments and provided a framework for analyzing transition dynamics.

As expected, long duration spells of extreme events are rare cases in our analysis. Using the 99th percentile threshold, long duration HFS (above 15 d duration) were observed in less than half of the studied catchments. In catchments where they occur, these spells are rare, occurring on average only once every 20 years in the most affected catchments. The spatial distribution of HFS with longest durations shows a concentration in northern France, consistently with regions characterized by chalk aquifer dynamics and by catchments with long memory (de Lavenne et al., 2022). In aquifer dominated systems, groundwater contributions can sustain elevated streamflow conditions over extended periods, eventually generating groundwater flooding, as mentioned previously.

Beyond detection, spell characterization can also provide insights for risk managers. In our study, it was shown that long duration HFS are not necessarily the most severe in terms of accumulated water volume. It was observed that streamflow in long duration HFS tends to remain above the threshold for a smaller percentage of time compared to medium duration HFS. This is particularly observed when using the higher thresholds (99th percentile compared to 95th), where even short duration HFS can occasionally exhibit higher accumulated water surplus than long duration HFS (Fig. 7). This was illustrated in the Mediterranean area (Rhône-Méditerranée hydro-administrative region, Fig. 3), where approximately 19 % of short and medium duration HFS had greater accumulated water surplus than the maximum observed in long duration HFS within the same catchment.

The classification of spell duration categories allows the identification of potentially more severe consequences in temporally compound settings, such as successive long duration spells or transitions from long duration LFS to medium or long duration HFS. These combinations are potentially more impactful than transitions between shorter duration spells. The analysis of transition between hydrological spells can reveal important patterns and enhance knowledge on compound events. In our study, we observed that transitions predominantly occur between short duration spells and within the same type of spells, either consecutive HFS or LFS in France. For example, within-a-month transitions between two short duration spells are more common than transitions from long duration LFS to short duration HFS, and transitions between two long duration spells are rarer. This frequency distribution reflects not only the prevalence of shorter hydrological spells, but also highlights the statistical challenge of analyzing rare, potentially high impact transitions between prolonged extreme conditions (Brunner et al., 2021).

We also investigated transitions from HFS to LFS. According to Hammond et al. (2025), sub-seasonal flood-to-drought transitions involve mechanisms that quickly deplete or fail to replenish surface and subsurface storage, such as early snowmelt, rain-on-snow events, or intense precipitation leading to high runoff partitioning. In our study, we identified that HFS to LFS transitions are predominantly slow, with the majority lasting over 90 d, while sub-seasonal within-a-month transitions are rare. However, for the rare cases where these rapid transitions did occur (in less than 14 d), our analysis indicates they were usually linked to antecedent low flow conditions in the previous two weeks, often acting as brief interruptions of a persistent LFS. This dynamic was detected in the 2003 drought, for instance, in a tributary of the Adour basin (see Fig. 3). Following a summer characterized by record high temperatures and an absence of precipitation leading to drought conditions in France, the region experienced two rapid transitions: from LFS to HFS, and returned back to LFS. The brief HFS interruption corresponds to intense precipitation events recorded in September 2003 (DIREN Midi-Pyrénées, 2003). These events likely generated high runoff partitioning that prevented significant recharge, causing the catchment to quickly revert to low flow conditions. We note that, in our study, the frequency of these rapid back-and-forth sequences may be underestimated at flashy catchments due to the use of daily mean streamflow data, which can smooth short duration flood peaks (Sect. 4.1). The use of hourly data, when available, could reveal additional cases of these compound transition patterns.

Moreover, we identified distinct patterns for the mean seasonality across different transition types. Within-a-month transitions showed a lower proportion of significant mean seasonality when compared to seasonal transitions, especially for consecutive HFS and transitions from LFS to HFS and HFS to LFS. Within-a-month and seasonal transitions from LFS to HFS predominantly occurred between October and December across most of France, with only a few high elevation catchments showing a preferential mean seasonality between February and May, mainly for seasonal transitions, located in the Alps and Pyrenees. This timing aligns with the broader European patterns identified by Brunner et al. (2025), who found that hydrological transitions from drought to floods generally occur in winter and spring across most of Europe, except in the Alps and Scandinavia where they tend to happen in summer. Consequently, the October-December predominance of within-a-month LFS to HFS transitions in our study likely reflects the typical French hydrological regimes with autumn/winter precipitation following the summer low flow periods. This timing differs slightly from the findings of Götte and Brunner (2024) in the Contiguous United States, where low elevation catchments showed a variable timing, although they similarly identified that transitions in high elevation, snowmelt-driven catchments tend to occur in summer.

Insights into spatial patterns can also be gained from the analysis of transition characteristics. In France, Mediterranean catchments showed very short LFS to HFS within-a-month transition durations, with median transitions of just eight days in the SPC Méditerranée Est (including Corsica) and SPC Méditerranée Ouest. Similar rapid transitions were observed in the SPC Grand Delta and SPC Alpes du Nord within the Rhône-Méditerranée hydro-administrative region, as well as in the Rhin-Meuse hydro-administrative region (SPC Rhin-Sarre and SPC Meuse-Moselle). These regional differences may be linked to distinct precipitation regimes, with Mediterranean regions experiencing more intense convective rainfall events with flashy behavior that can promptly terminate the low flow conditions (Vigoureux et al., 2024).