We analyzed physiographic and morphometric parameters at the level of the whole watershed and for buffer zones defined as the areas within 100 meters from the riverbank of the main river within each watershed (Table 1). This delineation is based on the Uganda National Environment (riverbanks, wetlands, and lakeshores) regulations, which mandate a 100-meter buffer zone from the riverbank to protect river ecosystems and mitigate environmental degradation (NEMA, 2019). For each pixel within the catchment and the buffer zone, the downward relief was analyzed to capture local elevation differences, representing the maximum elevation difference from the pixel to the lowest point in a 200m radius. The Topographic Position Index (TPI) was calculated to determine the relative position of the landscape with respect to peaks, slopes, and depressions. The Topographic Wetness Index (TWI) was determined to establish the flow accumulation and convergence, which offers a measure of how likely an area is to accumulate water and result in flooding. The upslope curvature was calculated to understand the flow dynamics. Land use and land cover classes were analyzed for both the watershed and the buffer zones by manually digitizing 2022 very high resolution images available in Google Earth. Lithology data were obtained from the geological map provided by the mapping department of the National Agricultural Research Organisation (NARO) in Uganda.

We conducted five field visits between November 2021 and December 2023 to the watersheds to assess their overall environmental status and collect key data on land use and land cover, population dynamics, landscape transformations and also historical natural hazards. These field assessments were also part of our previous study on the history of lands transformation and the occurrence of natural hazards (Kanyiginya et al., 2025), which provided a foundation for the present analysis on flash flood dynamics, especially in relation to landscape characteristics. We assessed the streambed material and measured stream bankfull widths and depths. Streambed material was assessed following a classification of rivers protocol by Rosgen, (1994). This classification is one of the most widely adopted methodologies for classifying rivers and streams due to its simplicity and robustness, as it integrates both qualitative and quantitative assessments (Allan et al., 2021; Corenblit et al., 2007). We took three river cross-section measurements, each at a different point (top, mid, and lower sections) along the river and averaged them (Xia et al., 2010). The measurements of river cross sections and width/depth ratios were measured at bankfull, which we hypothesized would have an influence on the frequency of flash floods according to the protocol by Rosgen (1994) and Xia et al. (2010).

Tipping bucket rain gauges are often used in meteorological stations, research, and environmental monitoring due to their reliability, simplicity, and ability to provide data at high frequency (Gray & Toucher, 2019). We used these tipping buckets to compile a two-year rainfall database from each watershed covering about the same period during which the river watcher collected the data (November 2021 to December 2023). In each watershed, we installed one tipping bucket rain gauge (Davis 6466M model) in November 2021. The rain gauges were installed at a location that was suitable in terms of safety for the equipment, either upstream or downstream of the watersheds (Fig. 1). The gauges provided high-resolution rainfall data for assessing key rainfall factors associated with flash flood occurrences. Tipping buckets record precise time data each time the bucket tips (after 0.2mm). This means that the timesteps between each tip are not equal. To allow for consistent time series analysis, we restructured the timesteps to 30 minutes throughout the recorded period and summed all rainfall (in mm through the number of tips) that fell within a 30-minute period.

Flash floods are characterized by their rapid onset after heavy or excessive rainfall in a short period. This period generally spans less than 6 hours (NWS, 2019). Hence, we characterized the intensity and duration of the rainfall (as the total sum within the rainfall event), for this six-hour period, before each recorded flash flood occurrence (Marchi et al., 2010). Antecedent rainfall conditions govern soil moisture fluctuations in the landscape. Higher moisture content within the soil has been shown to increase the potential for runoff generation (Grillakis et al., 2016; Islam et al., 2022). Studies have shown that antecedent rainfall over periods of 2 to 30 days can be significant in determining the resulting flash floods (Marchi et al., 2010; Upreti & Ojha, 2021). We therefore calculated cumulative rainfall over different periods (1,2,7, and 30 days), starting from six hours before the onset of flash flood and bankfull events.

We computed basic statistics of mean, median, standard deviation, and maximum rainfall, and the interquartile ranges for the annual rainfall data recorded across all watersheds. We generated box plots to visualize seasonal rainfall variability between watersheds. We then computed Pearson’s correlation coefficients (r) to identify relationships between the different watershed characteristics and the frequency of flash flood and bankfull events recorded in the two years (Stein et al., 2021). Additionally, the Mann-Whitney U test was employed to assess statistical differences between land use and land cover characteristics, and the occurrence or non-occurrence of flash floods.

We also paid particular attention to non-occurrence events, i.e., moments when rainfall is recorded but no flash flood was observed. The observation of non-occurrence is important when, for example, aiming to identify rainfall triggering thresholds (Bogaard & Greco, 2018; Ramos Filho et al., 2021). We used the 99th percentile of one-day cumulative rainfall from the two-year dataset as the threshold for intense rainfall in each watershed. Using one-day cumulative rainfall was advantageous in minimizing the inclusion of unrealistic non-flash flood events by focusing on daily rainfall extremes, which potentially lead to flash floods. Furthermore, this allowed us to evaluate the moments when heavy rainfall did not trigger flash floods (e.g. Ramos Filho et al., 2021).

We recruited eight river watchers, one for each watershed. The river watchers are local citizens living in the respective watersheds tasked to monitor water levels in a river and record information related to flash flood characteristics. Five of the eight watersheds studied were within the reach of the parishes that were monitored by the geo-observer network—a group of local residents who had previously participated in our study on the inventory of natural hazards (Kanyiginya et al., 2023). These geo-observers were assigned an additional task of monitoring river water levels, a designation we coined as "river watchers." Leveraging the geo-observers for this task was advantageous, as they were already experienced in data collection and familiar with reporting via a mobile app. For the remaining three watersheds, where no geo-observer network was present, we collaborated with local leadership to identify suitable candidates to take on the role of river watchers.

The eight river watchers were trained to collect data with a smartphone. They used KoBoToolbox, an open-source tool for mobile data collection (Nampa et al., 2020). A structured questionnaire that we specifically designed for this research was uploaded on KoboToolbox containing key questions on the rainfall characteristics and river level, i.e., whether it had rained or not at the monitoring site on the day of reporting, if it had rained on the day before the reporting day, and if the water level had exceeded the riverbanks (Appendix 1). We also trained the river watchers to take photos of water levels at a specified monitoring point in each watershed, clearly showing the water level in the river (Fig. 2) and the flooded surroundings in case of a flash flood. This ensured consistency and minimized variability in water level observations, allowing for accurate comparisons of water levels over time. The river watchers then send the recorded information (reports) via a mobile network to a central server for further analysis. Data collection was undertaken for two years from September 2021 to September 2023. The river watchers were tasked to monitor daily three of the eight watersheds, while the remaining five watersheds were monitored only when it rained at the monitoring site. Monitoring of the watersheds by the river watchers was done at random time during the day. Financial support was provided to the river watchers for transferring the collected data via an internet connection.

Flash floods were identified when river watchers reported that the water had surpassed the riverbanks, while bankfull events were noted when the water reached the identified bankfull features without overtopping the banks. The accuracy of these events was further validated by photos taken by the river watchers, providing visual confirmation of the reported events. However, establishing the exact timing of flash floods was challenging, as river watchers typically waited until the end of the rainfall event to move to the monitoring site and record observations. The time that the river watchers moved to the monitoring point to make the recording was not in our control. Nonetheless, we used the end of the rainfall event recorded by the rain gauge as a proxy for the timing of flash floods and bankfull events. The temporal characteristics of these events were then assessed in relation to watershed characteristics and rainfall measurements from the rain gauges to understand their correlations.

Watersheds have average slopes varying from 19° to 26°, with no major differences between the slopes for the entire watersheds and those for the buffer zones, except for W7 and W8, which show lower slopes in the buffer zones (Table 2 and Fig. 3a). The highest altitude was found in W4, which is part of the Muhabura volcano (Fig. 3c). The TWI values for the entire watershed are generally homogeneous across all watersheds, with a slightly higher value for W4 (Fig. 3e). Results of land use and land cover analysis indicate that most watersheds have about 50% of their land under cropland, with the highest percentage in W3 (Fig. 4). However, W4 and W7 have lower cropland cover, with only 18 and 9% respectively, although W7 has the highest proportion of cropland in the buffer zones (over 50%), mainly banana plantations (Fig. 5). The highest forest covers are observed in W4 and W7 at 60% and 40% respectively, while W2, W3, and W5 had the lowest ones (Table 2 and Fig. 4). The forest and tree cover are in most cases located on hilltops and hillslopes. The forest cover in the buffer zone is highest in W4 and W1 which have both tree cover and shrubs in the buffer zones of the river (Fig. 5). The buffer zones of W3, W5, W6, and W8 are covered by wetlands, which are semi-cultivated and partly covered with papyrus (Fig. 5). The highest proportion of built-up area is at 22% in W5 and the lowest in W1 at 2.7% (Fig. 4). The buffer zones of W6, W7 and W8 have a higher proportion of built areas as compared to other watersheds (Fig. 5). When considering the river channel, W1 has the highest width/depth ratio of 20.8, clearly higher than the other streams (Table 2). W3, W5, W7, and W8 are all characterized by silt/clay/mud as their dominant bed material. On the other hand, the bed material for the river in W1 is dominated by cobbles, and that of W6 is with sand and gravel.

Table 2 Watershed physiographic and morphometric characteristics. Lithology information is from the geological map of NARO (National Agricultural Research Organization), Uganda. Soil information is from the World Reference Base for Soil Resources (WRB 2014, Update 2015). https://files.isric.org/public/WRB/WRB2014_soil_map.zip accessed on 15/02/2025

The average annual rainfall received in the watersheds, without considering the three rain gauges for which the record was incomplete (Appendix 3), is 1598 mm for 2022 and 1190 mm for 2023, clearly showing that 2022 received higher rainfall than 2023 (Figs. 6a and b). The common seasonality patterns are present for the two years, although for 2022 we can note that August, considered as the last month of the dry season, was unusually wetter. The rainfall measured from Kabale station shows a similar pattern, with 2023 having lower annual rainfall than 2022. Note that the missing rainfall data are due to the malfunctioning of rain gauges, in W2, W3, and W8 (Appendix 3). For these watersheds, care was taken during rainfall analysis by considering only the months with full data. The extensive missing data in W2 necessitated removing this watershed from our analysis.

We generated a dataset of 1408 river observations (reports) from the data collected by the river watchers. W2 was excluded from comparative analysis due to data loss, as the river watcher lost the smartphone used for recording. Out of the data collected, we identified 20 flash floods and 10 bankfull events. Two flash floods were fatal events that, in total, resulted in l5 deaths, displaced several people and destroyed properties. For two flash floods and four bankfull events, the corresponding rainfall data associated with them was not recorded due to the malfunctioning of the rain gauge at the time.

We recorded the highest frequencies of flash floods and bankfull events in W3 and W8 (Fig. 4). Of the total recorded flash flood and bankfull events, 68% occurred in 2022 (Fig. 6). May and September experienced the highest frequencies of both flash floods and bankfull events, while no such events occurred in June, July, and August (Fig. 7). The recorded events coincided with the land preparation, planting, and weeding seasons, as outlined in the region’s farming calendar (Fig. 7). Additionally, we observed that more events occurred in watersheds that are intensively cultivated, particularly those with a high proportion of cropland (Fig. 4). A positive correlation (0.6) between cropland and the occurrence of flash floods and bankfull events is observed but is not statistically significant (P-value 0.15); meaning that the potential effect, if any, of cultivated areas on the influence of flash flood occurrence is to be considered with caution. Similarly, watersheds with less than 15% forest cover experienced relatively higher occurrences of flash floods and bankfull events compared to those with more forest cover (Fig. 4). This is associated with a negative correlation of − 0.6 (P-value 0.15), indicating that less forest cover could play a role on positively influencing the occurrence of flash floods. The same trend is observed when land use and land cover for the buffer zones are compared with the flash flood patterns, whereby buffer zones with less forest cover are inclined to higher flash flood frequencies (Fig. 5).

Analysis of the river buffer zones also shows that flash floods are experienced in watersheds whose buffer zones have wetlands, whereas watersheds without wetlands experienced non-flash flood events. The correlation between the occurrence of flash floods and wetland areas of the river buffer zone suggests a positive correlation of 0.7 (P-value 0.075) that must also be considered with care. Overall, although we observe some trends, their statistical significance is questionable.

The analysis of rainfall intensities higher than the 99 percentiles allowed us to identify 17 moments when flash flood where clearly absent despite high cumulative rainfall recorded in the rain gauges (Table 3). In our study, we coined these moments as ‘non-flash flood’ events.

The non-flash flood events relative to actual flash floods were high in W1, W4 and W7. On the contrary, the non-flash flood events were relatively less common in W3, W5, and W6 (Fig. 4). This pattern can be explained by the possible role of forest cover influencing the spatial distribution of flash floods; a negative correlation of − 0.6 (P-value 0.15) between the forest fraction coverage and the occurrence of non-flash flood events being observed, but not significant Whereas flash flood events were frequent in the wet seasons, the non-flash events occurred during both the dry and wet seasons (Fig. 7).

The majority (> 80%) of the flash flood and bankfull events occurred after rainfall exceeding 10 mm within 3–6 hours (Table 4, Fig. 8a). The cumulative rainfall associated with flash flood events shows a range of values extending from 4 mm to 53 mm (in 6 hours or less), with a median of 16 mm and a mean cumulative rainfall of 21 mm (Appendix 4). The fatal flash floods were triggered by a cumulative rainfall of more than 40 mm in 3–6 hours. For instance, the fatal flash flood which occurred on the 24th of January 2022 in W4 was triggered by a cumulative rainfall of 41 mm in three hours (Fig. 8a). The other fatal flash flood that occurred in W5 on 3rd March 2023 was caused by a cumulative rainfall of 53 mm in six hours (Fig. 8a, Table 4).

Generally, W3 and W8 experienced flash floods with relatively low rainfall intensities as compared to other watersheds (Fig. 8a). Similarly, bankfull events occurred with low rainfall intensities (Fig. 8a). However, the rainfall intensities associated with non-flash flood events were generally higher, with longer rainfall durations than those responsible for flash floods (Fig. 8a). In terms of rainfall duration, Fig. 8a shows that events in W3 and W5 are characterized by longer rainfall durations compared to other watersheds. A correlation analysis between the triggering rainfall duration and watershed size yielded a positive value of 0.5, though not significant. However, the triggering rainfall intensity and duration for flash floods/bankfull events showed a strong negative correlation (− 0.86) and was statistically significant (P-value of 0.03) (Fig. 8a).

Five flash flood events were associated with high antecedent rainfall conditions within two-seven days, that is, events 3 (44mm), 4 (60mm), 8 (31mm), 9 (62mm), and 17 (39mm) (Table 4). A negative correlation of − 0.4 (P-value 0.067) was observed between the six-hour triggering rainfall and the two-day antecedent rainfall (Fig. 8b), suggesting that antecedent rainfall is playing a rather limited role for the triggering rainfall with the highest values. Events 10, 11, 14, and 15 occurred with high triggering rainfall and were not influenced by antecedent conditions in the past seven days (Table 4, Fig. 8b). However, for occurrences where the triggering rainfall was low, high antecedent rainfall was present. For example, event 4 in W3 was triggered by a rainfall cumulative of only 4 mm in 5.5 hours (Table 4). However, this event was preceded by two heavy rain events that occurred within three days before its occurrence, with a cumulative rainfall of 60 mm. Overall, W3 experienced flash floods with the strongest influence of antecedent conditions. When antecedent rainfall conditions were evaluated for the non-flash flood events, results indicated multiple events combining high triggering events and high antecedent rainfall, although not leading to a flash flood (Fig. 8b).

The river watchers have collected information to understand flash floods. Their observations extended beyond merely recording the occurrence or absence of flash floods. For instance, they also documented detailed descriptions of water levels, including whether the water reached bankfull conditions, overtopped riverbanks, or the extent of flooding. Such detailed information, which is difficult to capture using conventional methods like satellite imagery or automated gauges, enhances the accuracy of flash flood assessments. The added value of citizen science approaches in understanding natural hazards has been highlighted by other studies globally (Hicks et al., 2019; Juang et al., 2019), and in a neighbouring region of Uganda (Jacobs et al. 2019; Sekajugo et al. 2022), underscoring its effectiveness in contributing to hazard monitoring where traditional data sources are limited or unavailable. This success highlights the potential for adopting this river watcher approach in other regions to enhance river monitoring and improve flash flood understanding, particularly in data-scarce areas.

Despite the positive results produced, the study encountered some limitations. With three watersheds monitored daily over two years regardless of rainfall, we should have expected more than 2,000 reports. In addition, we should add the reports from the other five watersheds that were monitored only when it rained. However, 1,408 reports were received, indicating that some river watchers did not contribute as expected. Discrepancies were most common among river watchers who lived far from the rivers, and some who faced technical challenges with their phones. For instance, the river watcher in W5 resided far from the monitoring site, which could have resulted in the non-recorded events that were expected to occur with heavy rainfall (Table 3). As reported by Ashepet et al. (2024) and Sekajugo et al. (2022), various factors motivate citizen scientists to efficiently report on events.

Another challenge was the timing of flash flood reports. Flash floods, by definition, occur rapidly in association with a rainfall event, but their exact onset time was not recorded because river watchers typically waited until the rain stopped to visit the monitoring sites. If a flash flood occurred at night, records were made the following day. Therefore, in our analysis, we considered by default the end of the rainfall event within six hours as the onset of flash flooding. The delayed reporting by the river watchers may have impacted the identification of a few bankfull event records, as water levels tend to recede over time. However, in most instances, the river left visible marks indicating where the water flow reached, which were corroborated by the photographs taken by the river watchers, providing a reasonable level of validation.

Our results suggest that watersheds with higher forest cover, both across the entire watershed and within the buffer zones of rivers (including W1, W4, and W7), appeared to experience fewer or less frequent flash floods and bankfull events. Forest cover is known to reduce the sensitivity to flash floods through, for example, increased rainfall interception, increased transpiration, and increased permeability of soils (Rogger et al., 2016; Sun et al. (2022); Hoang and Liou (2024).

Oppositely, the highest frequencies of flash floods appeared to be associated with watersheds (W3, W5, and W8), which had a large portion of cultivated land and the presence of wetlands (Figs. 4 and 5). The cultivated areas are associated with a runoff coefficient typically higher than what is observed in forest and grassland (Nyssen et al., 2004). Our results showed that a majority of the flash floods occurred in the land preparation and planting seasons according to the seasonal calendar of the region (Fig. 8). Planting seasons, when the ground is not protected by vegetation, are closely linked to an increase in surface runoff (Boardman et al., 2003). In addition, the cultivated areas in the Kigezi highlands are a result of relatively recent land use changes, whereby more vegetated land had been converted to cropland (Kanyiginya et al., 2025; Kilama Luwa et al., 2021; Twongyirwe et al., 2011). Land use change is clearly a factor that influences the incidence of flash flooding, especially in small catchments (Sun et al., 2022; Merz et al. 2021).

The morphology of the rivers could also influence the occurrence of flash floods. This is our assumption for W1, where the high width/depth ratio of the river channel (w/d = 20.8) might explain the absence of recorded flash floods in this watershed. Indeed, Yanites et al. (2010) explains that a width/depth ratio above 12 means reduced stream power and low erosivity, which reduces the possibility of flash floods occurring and bankfull conditions. On the other hand, the highest frequency of flooding was experienced in watersheds W3 and W8, where the rivers have the lowest ratio of 0.8 and 1.4, respectively. Such morphologies are associated with increased flash flood potential (Hadian et al., 2023). Added to the river's size is the streambed material in W1, which is composed of gravel and cobbles. Results revealed that flash floods were more frequent in watersheds that have bed materials consisting of silt/clay/mud. Large streambed particles tend to slow river flow velocity by increasing surface roughness and resistance, whereas small-sized particles create smoother streambeds that facilitate higher flow velocities, which can contribute to flash floods (Cohen et al., 2010).

Although lithology and soil characteristics influence flash flood patterns (Luu et al., 2023; Zenebe et al., 2013), our findings did not show a clear correlation between these factors and the occurrence or absence of flash floods. Instead, land use and land cover appeared to have a more pronounced impact on flash flood dynamics in the study area.

With a record over a period of two years for a relatively small number of watersheds, our dataset does not allow us to carry out strong statistical analysis as it can be done in regions with a tradition of long-term monitoring (Ávila et al., 2016). However, our field-based observations are detailed and reliable, and as such, bring a robust piece of information on the characterization of flash floods in a rural tropical environment of the Global South. Our measurements show that 2022 received higher rainfall than 2023, which may have contributed to the higher number of flash floods that we recorded that year. Although the data collected in this study is too limited to make definitive statement about the rainfall trends in the region, the potential link between flash flood occurrences and rainfall that seem to deviate from the average somehow offers a glimpse into the increasing incidence of flash floods we might expect in the future climate, where more intense rainfall is projected (Palmer et al., (2023). Fowler et al., (2021).

Our results also show that over the two years, the differences in monthly and annual rainfall totals vary between watersheds, although the topographic conditions are overall relatively similar (Table 2, Fig. 6a and b). This clearly shows the importance of the localized nature of intense rainfall that is characteristic of tropical climates, especially in mountainous regions (Dialynas & Bras, 2019; Zipser et al., 2006). For some of the flash flood events, especially the fatal ones, the association with high intensity triggering rainfall is evident. However, the results indicate that this is not always the case. The study identifies the occurrences of multiple other high intensity rainfall events that do not trigger flash floods (Appendix 2). Watershed (W1) for example, did not experience any flash floods while it received the highest daily and monthly rainfall extremes (Appendix 2). Conversely, some watersheds showed elevated river levels coinciding with low six-hourly cumulative rainfall, but these are then typically associated with higher antecedent rainfall over a 2–7-day time period (e.g. events in W3 and W8).

As confirmed by the analysis, the six-hour triggering rainfall for the non-flash flood events had a higher median than for flash flood events (Appendix 4), suggesting that factors beyond rainfall contribute to flash flooding. Similar findings were reported by Marjerison et al., (2016), who evaluated the influence of various variables on flash floods and found no significant correlation between the intensity of the triggering rainfall and flash flood patterns. This relationship was also observed in a study by NWS (2019), where reported flash floods did not align with the rainfall patterns of the area studied.

Antecedent rainfall conditions played a role in the occurrence of some flash flood events (Table 4). The influence of antecedent rainfall was more apparent for flash flood events where the triggering rainfall intensity was low. Nikolopoulos et al., (2011) and Zhai et al., (2018) demonstrated that moderate rainfall can generate significant runoff and contribute to flash flooding in previously saturated soil conditions. This is because soil that is saturated or nearly saturated from previous rainfall events has a limited capacity to absorb additional water, leading to rapid runoff and flash floods (Marchi et al., 2010; Merz & Blöschl, 2003). The influence of antecedent wetness is also common in wetland areas because of their high-water table, where even low-intensity rainfall can rapidly trigger flash floods. Our findings indicate that watersheds containing wetlands experienced both flash floods and bankfull events, whereas those without wetlands (W1, W4) were mainly characterised by non-flash flood events (Fig. 5). While antecedent catchment wetness can increase susceptibility to flash floods, some events in our study suggest it is not always a determining factor, as they occurred following relatively dry conditions (Table 4, Fig. 8b). This implies that when rainfall intensity is high enough, flash floods can still occur regardless of prior catchment wetness ( Archer & Fowler, 2015).

The observed mismatch between the cumulative amount of rainfall recorded and the occurrence or absence of flash floods can be attributed to the spatial differences between rain gauge locations, the flash flood monitoring points, and the upstream area contributing to the discharge at the point of monitoring. Significant differences in rainfall are observed sometimes over short distances as little as 100 meters (Table 3). While the distance between rain gauge stations and monitoring sites can contribute to discrepancies between recorded rainfall and observed hazard events (Papagiannaki et al., 2015), our study found no consistent pattern (Table 3). As shown in Fig. 9, the mismatch between rainfall recorded by rain gauges and observed river water levels can create misleading impressions, suggesting that flash floods can occur even under low rainfall conditions or that high-intensity rainfall does not necessarily trigger flash floods. These findings highlight the challenges posed by the localized variability of tropical convective rainfall (Archer & Fowler, 2015; Fowler et al., 2021) in the assessment of natural hazards.

The highly localised rainfall events not only affect the identification of triggering rainfall but also complicate the establishment of event thresholds (Nikolopoulos et al., (2014). Some of the rainfall threshold values identified in our study clearly do not reflect the actual conditions under which the flash floods were triggered. Nonetheless, the more reliable thresholds established in our study suggest values between 40 to 50 mm; less intense than the 80–100 mm thresholds reported by Monsieurs et al. (2018) in neighbouring regions. These differences could be related to the specific characteristics of our studied watersheds or to the fact that one of the two years observed, was wetter than the average, possibly leading to lower thresholds (Nikolopoulos et al., 2011; Zhai et al., 2018). However, due to limited observations, we cannot explore this further.

An interesting extra observation is that no landslides were observed during the occurrence of any flash flood or non-flash flood events in our study. In contrast, Monsieurs et al. (2018), further supported by Deijns et al., (2024; 2025), documented frequent landslides occurring alongside flash floods. Such compounding flash flood- landslides events are associated with high rainfall intensities. The absence of landslides in our study might be indirect evidence that provides further support to the lower threshold hypothesis. In other words, the flash floods that we observed were not associated with very extreme rainfall conditions, which is something to be expected over a limited period of observation on a rather small region. These flash floods somehow represent the everyday hazard pattern. They do not get a lot of attention and do not make the highlights, but the cumulative “small” impacts of these extensive risks (UNDRR, 2017) can lead to a accumulation of problems that outweigh those associated with the occurrence of extreme events (Bull-Kamanga et al., 2003).