Aiming to detect modern, natural changes resulting from mass-wasting processes along the studied escarpment, we compared aerial photos and high-resolution digital surface models that were available for the years 2013–2019. During that period, no additional instrumentation was installed in the study area to identify triggered DFs. We therefore used the available orthophotos and digital surface models to map new DF deposits and to minimize the time interval of triggering. The digital surface models are the product of airborne light detection and ranging (lidar) scans, having a spatial resolution of 0.5 m per pixel with spatial uncertainty lower than 1 m. In the vertical dimension, the digital surface models' absolute elevation error is lower than 0.35 m. We obtained digital surface models for every second year: 2013 (scan date 29–31 May 2013), 2015 (April–May 2015), 2017 (12 May 2017) and 2019 (29–30 September 2019). In order to identify topography changes, we subtract the earlier digital surface model from the more recent one. The new raster of elevation differences, also known as digital elevation model of difference (DoD), should have positive values if material added, negative for deficiency and around zero for no change (Williams, 2012; Morino et al., 2019). In general, this is the case apart from two exceptions: (1) the average difference value for stationary areas is non-zero, suggesting a vertical offset of a few tens of centimeters; and (2) noisy results around high relief lines (cliffs). The first issue, probably caused by a systematic error in the digital surface model elevation, is expected to have a minor influence on our results: to ensure a better detection we computed the mean offset by averaging several stationary areas and subtracted it from the DoD. The second issue mostly affects the cliff area, and the cliff lineaments in particular, and consists of noisy raster values (neighboring cells with values above and below zero). This artifact is not expected to affect our detection because DFs are recognized by nearby regions with similar values, so noisy areas can be easily excluded. However, some tiny DF deposits are possibly missed by this procedure, especially those with an area smaller than 25 m2 and average thickness smaller than 0.6 m.

In order to narrow our uncertainty in the estimated time of triggering (otherwise only limited by the time interval between airborne lidar scans), we used orthophotos, media and social network documentation. In terms of aerial photos, we mostly use geometrically corrected orthophotos with a resolution of 0.25 m per pixel from the years 2015 (photos taken in June–July 2015) and 2016 (photos taken 15–30 April 2016). Nevertheless, for most cases, the exact timing of the DF triggering was unknown, and we had to identify the most probable triggering date as described below.

We identified 43 DF deposits that occurred between the years 2013–2019. Two additional deposits were classified as rockfalls and removed from the analyses. The DF deposits are located along small ephemeral streams that drain the cliff area above them and present evidence of previous DF activity. This suggests that they were mobilized by flow events and therefore considered DFs, as suggested by Hungr et al. (2001). The DFs we examine are small-sized, with a maximum runout distance (between the source location and deposit) of a few tens of meters. The areal extent of each DF deposit ranges between 10 and 1000 m2 with an average value of 150 m2 and elongated shapes of about 30 m in length and 5 m in width. Although during the studied period (2013–2019) only small-sized and short-runout-distance DFs could be clearly mapped, longer-runout-distance DFs from past events are observed along the studied escarpment (Fig. 3). In principle, it is possible that larger DFs did occur during the studied period (2013–2019), but subsequent road construction and floods blurred the DF tracks. This made it impossible for us to detect them by means of DoDs and to verify the nature of the phenomena (e.g., distinguishing between DFs and flash floods) based on field surveys. This is to emphasize that, although the studied DFs are relatively small, their triggering mechanism is similar to potentially more hazardous DFs. All the mapped DFs in this study are of the first to second stream order, following the “top down” system division. Their drainage basins are extremely small, usually limited to the cliff itself or to a kilometer west of the clifftop at most. The spatial extent of each DF deposit was mapped into a polygon based on the DoDs (positive values for added material). These polygons were primarily grouped according to the time interval they were triggered (2-year intervals based on the lidar scans): 17 DFs between 2013 and 2015 and 26 DFs between 2015 and 2017, while no deposits were identified for the period 2017–2019 (Fig. 1).

Aiming to verify that the mapped deposits are indeed the result of DFs, we explored their structure and morphology in the field (Fig. 4). In many cases, the DFs are located on very steep slopes (20–40∘) with limited access. All mapped deposits and their related upstream erosional scours (i.e., depressions along the streams where the missing material is the source for the debris material) are located along or at the edge of short ephemeral streams that drain the cliff. The deposited sediments are poorly sorted, a principal characteristic that distinguished DFs from water-laid sediments. Distal lobes, also known as depositional lobes, are observed for all the mapped deposits at the toe of the flow. For some cases, also side lobes are observed upstream along the flow path (Fig. 4e and f). Levees, usually a few tens of centimeters high, are occasionally observed on both sides of the channel, although they are barely seen in the DoDs. In the field, the greyish levees are clearly distinguished from the surrounding brownish colluvium (Figs. 3 and 4d). These levees consist of relatively large fragments, similar to those found in the lobes. All these elements suggest that these deposits were caused by DFs.

For most of the mapped DFs (37 cases), the deposits consist of unsorted dolomite and limestone angular fragments with a maximum size of a few tens of centimeters. This deposit composition reflects the colluvium composition above it, which is mostly composed of the upper-escarpment cliffs and some eolian fine particles. Hence, the source for the debris material is the talus at the base of these cliffs (Fig. 2b). In contrast, in fewer DFs (six cases) located at the bottom of the lower step on the escarpment, the deposits consisted of rounded fragments, apparently derived from the nearby cliff of lacustrine and conglomerate sediments (Fig. 2c). These same two sources of debris materials were also reported by a previous study in the region (Ben David-Novak et al., 2004).

We divided the mapped DFs into four groups based on the triggering time interval and on their spatial distribution across the study area, with the assumption that each group has been triggered by an individual convective cell (Fig. 1). While it is in theory possible that nearby DFs were triggered by different storms over a short period of time, we deem this possibility highly unlikely due to the rare occurrence of DFs in the region. Conversely, a specific storm that may last for a few days, could represent the trigger of several groups. However, since the distances between groups observed in the same 2-year time intervals are always greater than 8 km and the typical scale of convective cells in the region is smaller (Belachsen et al., 2017; Marra and Morin, 2018), these groups were likely triggered by different convective cells. As we will see, this is possibly the case of our groups 1 and 4, which occurred in the northern and southern parts of the study area, respectively (Fig. 1).

Our mapping of natural changes along the steep escarpment between the years 2013–2019 suggests that the majority of mass-wasting events were caused by DFs. We used both rainfall stations located within or in close proximity to the study area and the weather radar of the Israel Meteorological Service (Fig. 1) to identify and characterize the triggering storms. We used the rainfall data from two types of stations (https://ims.gov.il/en/stations, last access: 6 March 2023) as shown in Fig. 1: (1) automatic rain gauges with a time resolution of 10 min (from north to south: Bet Haarava, Metsoke Dragot, En Gedi bath); (2) daily rain gauges (from north to south: Almog plantation, Enot Zuqim, Mitzpe Shalem factory). Weather radar data are used to assess the rainfall conditions above the DFs, and rain gauge data are used to constrain and validate the radar data on a regional scale.

The weather radar (https://ims.gov.il/en/node/193, last access: 6 March 2023) gives rainfall intensity maps every 5 min at 500 m resolution, and allows for the spatial analysis of the event. In particular, weather radar data are essential to capture the DF-triggering rainfall conditions especially in convective environments. In fact, rain gauges tend to systematically underestimate the DF-triggering rainfall because they sample precipitation fields away from the triggering locations and because these fields are characterized by sharp spatial variability and necessarily have high intensities over the triggering locations (Marra et al., 2016, 2014; Destro et al., 2017; Nikolopoulos et al., 2014, 2015). Weather radar data were provided by the Israel Meteorological Service and elaborated following the procedures described in Marra et al. (2022), which include physically based corrections and empirical adjustments based on rain gauge climatology. In addition, in order to reduce the bias typically observed in our study area (e.g., see Marra et al., 2022), we gauge-adjusted the radar data of each storm (see definition above) using data from the daily rain gauges described above (e.g., see Rinat et al., 2021). Despite these careful adjustments, the radar samples the study area about 3 km above the ground (see Marra et al., 2022), so rainfall at the ground could be misplaced by a few hundreds of meters in the case of strong winds in the lower atmosphere. This may cause some errors in the estimation of rain intensities; in particular, because of the typical characteristics of DF-triggering convective rainfall (Marra et al., 2016), we expect that in these cases, our evaluation could be underestimated. While this does not affect our findings, some caution should be taken regarding the rain intensity values we report. It is worth noting that the previous studies in the region (e.g., Ben David-Novak et al., 2004) were based on data from a different weather radar, which was, however, located not far from the one used here; thus, the same caveats are to be used in interpreting those quantitative estimates.

Since the study area was documented by airborne lidar roughly every second year, we searched for all the potential triggering storms that could have triggered the DFs during the time intervals between two subsequent scans (Fig. 6). For the analysis, we consider the location of each DF at the center of its mapped deposit. The many streams that drain the cliff have a very small drainage area (maximum distance between the deposits and the basin head is 750 m). As the weather radar resolution is 500 × 500 m2, we can assume that the rainfall observed by the radar corresponding to the deposit location represents a good approximation of the rainfall over the drainage basin (Marra et al., 2014).

Following this procedure, we plot for each DF the 5 min rainfall intensity versus time for its relevant 2-year interval (Fig. 6). For example, for a DF from group number 3, mapped east to Metsoke Dragot, we used the radar data for the years 2013–2015. Note that gaps in the radar data are possible because of technical issues and because during dry periods the radar is often turned off. In order to make sure that we did not miss critical storms, we compared the radar data with rain gauge stations. During the period of 2013–2017, we found only two storms (8–9 January 2016 and 27–29 January 2017) during which the radar was off.

We picked potential triggering storms by isolating those storms in which the peak rainfall intensities exceeded 10 mmh-1 and the peak 30 min intensity (average intensity over 30 min periods) exceeded 5 mmh-1 (Fig. 6). Recalling that previous studies reported thresholds of 30 mmh-1 over 1 h periods and that 10 mmh-1 is often regarded as a threshold for defining convective precipitation (e.g., Peleg and Morin, 2012), this selection is based on rather low intensities; i.e., it ensures we will select all the potential triggering storms, and it diminishes the possible impact of radar underestimation mentioned above. It is worth noting that this selection is only based on intensity, and no other condition on rainfall amounts or rainfall preceding the high-intensity is used.

In total, we identified 11 potential triggering storms that satisfy the above intensity conditions over the DFs during the periods of interest (Table 1). Nevertheless, only 7 of these 11 storms satisfied the conditions over all the mapped DFs of a group of interest (marked by bold text in Table 1 and blue rectangles in Fig. 6). Therefore, only seven storms can be considered potential triggering storms. Note that one of them (26–29 October 2015) is a potential triggering storm for the DFs of two groups. We should remember that the process of identifying the triggering cell could be subject to errors, so a different storm or cell or more than a single convective cell may have triggered the mapped DFs in a group. The procedure we adopt, however, closely follows our knowledge of DF triggering in general and of local climatology of storms in particular, so this possibility is expected to be unlikely.

Once we identified the potential triggering storms, we needed to narrow down our choice to the most probable storm that triggered each DF group. To achieve that, we used additional analysis of the radar data, orthophotos, and online reports from social media and news websites. For each of the seven potential triggering storms we examined the event time series in terms of total rainfall and rainfall intensity versus time for a period of hours to a few days (Fig. 7). We also produced the 5 min resolution rain intensity maps such as the one shown in Fig. 8. Using those, we followed the candidate convective cells of each storm in time and examined their spatiotemporal behavior with respect to (i) the mapped deposits and (ii) areas with similar slopes and/or sediment availability but no observed DF. This allowed us to narrow our choice. In some cases, such as group 1 in the northern part of the study area, we could not clearly pinpoint a single triggering cell for the DFs in the group, and we needed additional information. Using two orthophotos from the years 2015–2016 we limited our search for two out of three storms: 26–27 October 2015 and 10–13 April 2016 (Table 1). We manually searched several websites (including Facebook, YouTube and newspapers) for relevant photos and videos of floods that could help us determine the flow characteristics (debris flow or flash flood) and identify the exact timing of the event. We could only find a Facebook video showing deposits on the main road (route no. 90) (https://www.facebook.com/chen.sason.75/videos/1155321937831332, last access: 6 March 2023) that crosses the most southern stream of mapped deposits from group 1 on 27 October 2015. We could use this information to pinpoint a specific event and a most likely triggering cell for group 1 DFs: 27 October 2015 between 13:55–14:40 LT (UTC+2). Combining the above information, we ended up identifying the four triggering cells that, based on the information at hand, most likely triggered the DFs in the four groups (Table 1).

The characteristics of the rainfall during the triggering storms were examined to evaluate the critical conditions for DF triggering in the study area. The triggering cells are intense and short: intensities are in the range of 10–85 mmh-1 and durations between about 20 min (groups 2–4) and 45 min (group 1; Table 1 and Fig. 7). Interestingly, for all DFs, significant antecedent rainfall was observed for the triggering storms during the hours preceding the triggering cell (usually more than 10 mm; Fig. 7). To further explore this point, we plotted for each group the spatial distribution of antecedent rainfall over the study area and compared it with the spatial distribution of the total rainfall yield of the triggering cells, as well as with their maximum intensities over 5 and 30 min intervals (an example of such a plot is given in Fig. 9). Figure 9b–d present the triggering cell characteristics of group 3 (in the center of the map) and suggest that two additional convective cells (warmer colors) north and south to the main cell passed through the studied area at the same time. While the maximum intensities of all three cells exceed 10 mmh-1, only the central one triggered the DFs (of group 3). Although the southern cell covers the steep slopes well (meet the steep slope and sediment conditions) and is similar in rainfall intensity, no DF was triggered. We suggest that this is related to lower amounts of antecedent rainfall. In contrast, for the northern cell it seems that the antecedent rainfall condition is available, but the cell only marginally crossed the steep slopes (the cell at its peak intensity did not cross the cliff area).

In order to validate the hypothesis that antecedent rainfall is an important factor in DF triggering in the study area, we examined the whole period between 2013 and 2019 with the aim of identifying storms which could potentially trigger DFs in terms of rain intensity but did not. We focused on the triggering locations of our DFs in order to ensure the susceptibility conditions are met. In Fig. 10 we plot the antecedent rainfall versus the 5 min rain intensity for a representative DF deposit from each group. Here, the antecedent rainfall is calculated automatically over the entire time series as all the rain accumulated on the deposit pixel over the 120 h preceding the specific measured intensity. This automatic definition provides estimates of the antecedent rainfall that are equivalent to the ones used earlier, except for the rare storms lasting longer than 5 d prior to the peak intensity; in these cases, this automatic definition might underestimate rainfall. For most cases, in these new graphs we picked the same potential triggering events as we previously showed (Table 1). Interestingly, for extreme-intensity rainfall events (exceeding 60 mmh-1), like the 15 September 2015 event in group 2 (Fig. 10b) and events on 4 June 2018 and 29 October 2015 in group 3 (Fig. 10c), no DF deposits were observed even after re-checking the target area in our DoD.

An interesting additional story is told by Fig. 10a: the event on 8 May 2014 shows relatively high antecedent rainfall (more than 10 mm), as well as peak intensity (more than 50 mmh-1), compared to the likely triggering event on 27 October 2015 but occurred before the DFs in group 1 were triggered (they were not observed in the lidar flight between these two storms). Given the strong characteristics of this cell in terms of both antecedent rainfall and peak intensity and since the mapping could not be fully effective for small DFs, we further examined the region around group 1 looking for potential signatures of the 8 May 2014 storm. Indeed, some tiny shallow deposits appeared during the period 2013–2015. They are located on the western edge of the group 1 area in close proximity to the 8 May 2014 cell peak. While these mass movements are small with respect to the DFs examined so far in this paper, they show similar DF-like properties. The fact that we could find new mass movements based on the requirement of both antecedent rainfall and peak intensity constitutes an additional element supporting our hypothesis.

As mentioned in Sect. 4.3, the investigated triggering cells are intense (10–85 mmh-1) and short (20–45 min) (Table 1 and Fig. 7). Both their intensities and durations are lower than the previously suggested threshold for DF triggering in the study area (intensities greater than 30 mmh-1 for duration of 1 h or longer; Ben David-Novak et al., 2004). A previous study conducted on the arid slopes of the Grand Canyon, Arizona, cautiously suggested that sustained intensity exceeding 20 mmh-1 and a total rainfall of 25 to 50 mm may be a minimum requirement for DF triggering, without mentioning a minimum duration (Melis et al., 1995). In our study, only two out of the four triggering cells have maximum intensities exceeding 30 mmh-1, but these intensities were observed only for a short period (less than 10 min), much shorter than what was previously reported (Fig. 7). While for most of our mapped DFs, intensity was indeed greater than 20 mmh-1, the total rainfall for all triggering cells was lower than 25 mm (Fig. 7). Quantitative accuracy of radar data, however, is not perfect, and, as mentioned above, possible underestimation cannot be excluded. Considering the limited datasets available for arid regions, it is still impossible to determine a unique threshold for DF triggering. More attempts should focus on data collection in these regions and on carefully considering the spatiotemporal distribution of rainfall during the triggering storms. Hints towards the importance of the temporal rain distribution during a storm could already be found in Ben David-Novak et al. (2004), as their data showed that for both studied events the triggering cells reached the area only hours after a significant antecedent rainfall.

We speculate that antecedent rainfall or the flow of water in the steep channels caused by antecedent rainfall could reduce the sediment strength by wetting. The effects of pore pressure increase and the reduction in apparent cohesive strength most likely cause this strength reduction (Takahashi, 2014; Terzaghi, 1943). The strength reduction may be followed by enhanced incision or direct failure of the sediments within the channel. Such slope instability can initiate the DFs once the high intensities provided by a triggering cell cause a significant water flow at these channels. Shmilovitz et al. (2020) showed that the threshold for runoff flow on nearby desert slopes is 14–22 mmh-1 for a duration of 5 min. This threshold is met by all the four triggering cells that we consider the most likely trigger for the mapped DF deposits, but it is also met by many other storms among the ones we identify as possible candidates (Table 1). We hypothesize that, due to the lack of antecedent rainfall, no DFs were triggered by these cells. At the same time, however, flash floods could indeed have been generated, as frequently observed along the cliff during heavy storms (Belachsen et al., 2017).

The need for antecedent rainfall could also help explain the low incidence of DFs in the area. To date, only seven modern DF-triggering storms are reported in the study area: four events reported for the 10-year period between October 1987 and October 1997 (Ben David-Novak, 1998; Ben David-Novak et al., 2004) and three identified in the current study – roughly equivalent to one triggering storm every 2–3 years. Given the vast sediment availability, the rarity of DFs in the area is usually explained by the dry weather (i.e., low number of storms) and the small areal extent of the susceptible steep slopes: only convective cells hitting the small susceptible area can trigger a DF. However, while storms in the area are indeed not frequent, each storm usually brings numerous convective cells. Belachsen et al. (2017) identified 424 storms in the area in the period 1990–2014, about 20 per year, associated with over 10 000 convective cells, an average of over 24 cells per storm. Similarly, Marra and Morin (2018) showed that individual storms can bring even hundreds of high-intensity convective cells. With these numbers, it is more difficult to explain the rarity of DFs just using the reasoning above. Conversely, the here-suggested need for antecedent rainfall as a prerequisite for the triggering implies a need for an intense convective cell to hit an area where non-negligible precipitation already occurred during the same storm.

The need for antecedent rainfall as a critical triggering condition also in arid areas could facilitate the prediction of such events. Heavy rain intensity remains a key trigger of DFs, but the trajectory and intensity of convective cells cannot be forecasted with sufficient accuracy even with the most advanced weather models (e.g., see Rinat et al., 2021). In addition, the short distance between the debris sources and the vulnerable structures drastically reduces the effectiveness of warnings based on radar nowcasting (e.g., Sideris et al., 2020). Conversely, antecedent rainfall can be monitored more easily during the storm, for example using weather radar observations, and then used to highlight areas more susceptible to possible incoming convective cells. This information could be used to send targeted warnings to the relevant locations and structures at risk.

Our results are based on a relatively small sample of small-sized DFs detected in an arid region of the eastern Mediterranean. Although we extensively explored the region in Fig. 1, this inventory cannot be considered complete because DFs could have been missed due to noise in the lidar data or other data issues (e.g., see Sect. 4.3). In addition, weather radar uncertainties may affect the precipitation estimates (e.g., see Marra et al., 2022). One source of uncertainty in particular is worth mentioning: the possible advection of precipitation during its falling from the height of the radar sampling volume (about 3 km above the ground level) to the ground could lead to the misplacement of the radar estimates by a few hundreds of meters; this would typically lead to an underestimation of the rainfall amounts (Marra et al., 2016). In addition, it is important to recall that the adjustment of the radar data is based on only a few available stations: some level of uncertainty in the quantitative estimates is thus to be expected. While our qualitative results are robust with respect to these sources of error and support our reasoning in terms of process description and understanding, the numbers are subject to residual uncertainty, and caution is advised against their direct use in warning systems.