Reconstructing the Salgar 2015 Flash Flood Using Radar Retrievals and a Conceptual Modeling Framework: A Basis for a Better Flood Generating Mechanisms Discrimination

Flash floods associated with severe precipitation events are highly destructive, often resulting in significant human and economic losses. Due to their nature, flash floods tend to occur in medium to small basins located within complex high mountainous regions. In the Colombian Andean region these basins are common, with the aggravating factor that the vulnerability is considerably high due to the presence of important human settlements frequently occupying floodplains and other 5 flash-flood prone areas. During May 18 of 2015, severe rainfall generated a flash flood in the municipality of Salgar, in La Liboriana basin located in the northwestern Colombian Andes, resulting in more than 100 human casualties and significant economic losses. This work aims to reconstruct the 2015 La Liboriana flash flood to understand better the hydrological processes that took place during the event and to serve as a proof of concept for the development of a flash flood guidance low-cost 10 tool when information is scarce. Radar QPE, satellite data, and post-event field visits are used to reconstruct the Salgar flash flood addressing the relationship between rainfall spatio-temporal structure, soil moisture, and runoff generation during successive rainfall events, using conceptual modeling framework including land-slide and hydraulic sub-models. The hydrological model includes the addition of virtual tracers to explore the role of runoff and subsurface flow, as well as the relative 15 importance of convective and stratiform precipitation in flash flood generation. The assessment of the interactions between runoff, subsurface flow, and convective-stratiform rainfall allows a better understanding of the short-term hydrological mechanisms leading to the flash flood event. The overall methodology reproduces considerably well the magnitude and timing of La Liboriana flash flood discharge peak, as well as the areas of regional land-slide occurrence and flood spots location, with 20 some limitations due to the available spatial scale of the digital elevation model. Simulation results 1 Nicolás Velásquez 1,2, Carlos D. Hoyos 1,2, Jaime I. Vélez 1, and Esneider Zapata 2 Hydrol. Earth Syst. Sci. Discuss., https://doi.org/10.5194/hess-2018-452 Manuscript under review for journal Hydrol. Earth Syst. Sci. Discussion started: 25 September 2018 c © Author(s) 2018. CC BY 4.0 License.


Introduction
Flash floods are regarded as one of the most destructive hydrological hazards resulting in considerable loss of human life and high costs due to infrastructure damages (Roux et al., 2011;Gruntfest and 35 Handmer, 2001).Flash floods are usually described as rapidly rising water level events happening in streams and rivers, associated with short-term, very intense convective precipitation systems, or orographically forced rainfall events over highly saturated land surfaces and steep terrains (Šálek et al., 2006;Llasat et al., 2016;Douinot et al., 2016).Convective precipitation episodes often feature high intensities, short durations, and relatively reduced spatial coverage (Houze, 2004).Jonkman (2005), 40 based on information from the International Disaster Database, maintained by the United States Office for Foreign Disaster Assistance and the Center for Research on the Epidemiology of Disasters in Brussels, shows that between 1975 and 2001 a total of 1816 worldwide freshwater flood events killed over 175,000 persons and affected more than 2.2 billion people.These events not only caused human and economic losses but also damages to ecosystems and loss of historical and cultural values.45 Among all different types of floods, Jonkman (2005) shows that flash floods result in the highest average mortality rate per event (3.62%), almost ten times larger than the mortality rate for river floods.
Several studies assess the role of rainfall, geomorphological features and soil conditions on flash floods occurrence from different points of view.Wu and Sidle (1995) highlight the role of the to-50 pography, ground cover, and groundwater in the occurrence of shallow landslides and associated debris flows.Merz and Blöschl (2003) highlighted the importance of understanding the physical processes giving rise to floods of a certain probability of occurrence and proposed a framework, from a catchment perspective, for identifying causative mechanisms of inundations and the viability for a large number of episodes and basins.Given the critical role of precipitation, some authors 55 follow a climatological approximation to assess the recurrence of flash floods in particular regions Due to the rapid nature of flash floods, they are more likely to occur in small and steep basins (Younis et al., 2008), highlighting the need for the availability of detailed temporal scale precipitation data (Norbiato et al., 2008) and high-resolution terrain information.Watersheds prone to these events are usually located in rural mountainous areas with poor gauge information (rain and level).
The utilization of radar data to study flash flood generating storms is vital for understanding and forecasting these events (National Research Council 1996).One of the most critical challenges associated with flash floods is their simulation and prediction (Yamanaka and Ma, 2017;Borga et al., 2011;Marra et al., 2017).The fact that flash floods are more prone in small basins increases their intrinsic physical and measurement uncertainty (Wagener et al., 2007), difficulting their prediction (Hardy et al., 2016;Ruiz-Villanueva et al., 2013).Many authors have assessed the influence of hills and stream slopes and evidence suggest the slopes of the hills are significantly more important for flash floods occurrence and magnitude than the stream slope (, eds.; Šálek et al., 2006;Roux et al., 2011;Yatheendradas et al., 2008).
Arguably one of the most meaningful applications of quantitative precipitation estimation (QPE) from a societal point of view is its use to improve flash flood forecasting.The use of weather radar QPE products in flash flood applications could potentially offset, in part, the lack of in-situ precipitation products in small poorly gauged basins and become an important stepping stone for improving the state-of-the-art understanding of flash flood processes such as orographically enhanced extreme rainfall events and runoff generation (Creutin and Borga, 2003a).However, in some cases, there still exist limitations associated with the spatial resolution of radar retrievals (Šálek et al., 2006;Hardy et al., 2016).Nevertheless, several modern approaches involving radar information, in-situ precipitation data, model approximations and merging algorithms are promising (Braud et al., 2016).
The topography of Colombia is characterized by three branches of the Andes crossing the country south-to-north, generating a mixture of landscapes from high snow-capped mountains, vast highland plateaus, deep canyons to wide valleys, making some regions highly prone to flash flood occurrence.
The likelihood of flash flood occurrence in Colombia is also high due to the spatio-temporal behavior of the Intertropical Convergence Zone, and the direction of the near-surface moist air flow leading to orographic enhancement of convective cores (Poveda et al., 2007).In the last decade, there have been several widespread and localized flash flood events in Colombia associated with climatological features and with the local intensification of rainfall events.The 2010-2011 La Niña event alone triggered 1233 flooding events and 778 mass removal processes, with more than 3 million people affected and damages estimated by the "Comisión Económica para América Latina y el Caribe " in more than 6.5 billion US dollars.After this widespread disaster, several localized events have occurred in the country with devastating consequences.covered almost all the basin (hereafter referred to as event 1).Twenty hours later, between 23:00 P.M of May 17 and 02:00 A.M. of May 18 two successive extreme convective systems occurred over the basin with the maximum intensity in the upper hills (event 2).Event 1 corresponds mainly to a stratiform event, covering almost all the basin area and with an average precipitation accumulation of 47 mm.Event 2 accumulated, averaging over the basin, around 38 mm, however over the upper 140 watershed the accumulation exceeded 180 mm.The occurrence of these successive rainfall events resulted in a flash flood that took more than 100 human lives and affected several buildings and critical infrastructure.The total investment for the reconstruction of Salgar´s infrastructure and the psychosocial attention of the community is estimated at 36.000 million Colombian pesos (approximately 12,5 million dollars considering 2018 exchange rate), which corresponds to three times the 145 annual income of the municipality.Figure 1 shows an example of infrastructure damage as a result of the flash flood event, and satellite images of the basin's main channel before and after the flash flood, showing considerable river margins and bed erosion.
In this case, there are no local records of soils or land use and the local hydro-meteorological data is scarce or non-existent and certainly not available in real time.Due to the lack of data, La Liboriana case imposes a challenge for flash flood prediction and modeling.According to Blöschl et al. (2012), there are three methods for using models in PUB cases.The first strategy is to obtain the required 155 model parameters from historical basin behavior and the morphological characteristics of the basin.This strategy often leads to low model performance (Duan et al., 2006).The second approach is to inherit the hydrological calibration from a gauged neighboring watershed, which in this case does not exist.The third method is to parameterize the model based on proxy variables, such as hydraulic information obtained during field recognition.In the case of the 2015 La Liboriana basin flash flood, 160 there are no previous historical streamflow records, nor records from a neighboring watershed.
The present work aims to reconstruct the 2015 La Liboriana flash flood, not only to understand better the hydrological processes that took place during the event but also to serve as a proof of concept that hopefully leads to the development of a flash flood guidance low-cost tool based on the 165 scarce information existent.Such a tool should be useful for decision-makers not only for short-term decisions in the context of an early warning system but also as a planning resource for long-term risk management.In particular, this work uses information derived from radar QPE, satellite data, and post-event field visits to reconstruct the Salgar flash flood event.Also, this study addresses two broad hydrological issues.The first issue consists in exploring the relationship between rainfall 170 spatio-temporal structure (Llasat et al., 2016;Fragoso et al., 2012), soil moisture and runoff generation (Penna et al., 2011;Tramblay et al., 2012;Garambois et al., 2013) during the successive rainfall events.The second issue consists in proposing a simplified hydrological modeling scheme including land-slide and hydraulic sub-models to detect and anticipate potential flash flood events.

175
The methodology followed in this study makes use of a conceptual modeling framework that  Vélez (2001) and Francés et al. (2007), simulates different hydrological processes as independent but interacting storages.The shallow landslide sub-model follows the formulation described in (Aristizábal et al., 2016).The hydraulic sub-model corresponds 180 to a low-cost 1D model (hereafter referred to as HydroFlash) that assumes infinite sediment supply and estimates the cross-sectional filled area at all time steps.The hydrological model includes the addition of virtual tracers to explore separately the role of runoff and subsurface flow, as well as the relative importance of convective and stratiform precipitation in flash flood generation.The convective and stratiform classification follows the methodology proposed by Houze et al. (2015) 185 and Steiner et al. (1995).The assessment of the interactions between runoff, subsurface flow, and convective-stratiform rainfall allows a better understanding of the short-term hydrological mechanisms leading to the flash flood event.The model is parameterized using field information from a river cross-section located at the entrance of the urban area of Salgar.A comparison between the results from both sub-models and the observed landslides scars and flooded spots helps to evaluate Section 4 also includes a discussion on the role of the rainfall structure in the flash flood reconstruction and presents results from the land-slide and inundation sub-models.Finally, the conclusions are 200 presented in section 6.

Data and Area of Study
The urban area of Salgar municipality is located near the outlet of La Liboriana basin, a small (56 km 2 ) tropical watershed located in the westernmost range of Colombia's Andes (Figure 2).La Liboriana basin joins El Barroso river basin and both drain to Cauca river.The availability of the 205 ALOS-PALSAR Digital Elevation Model (DEM) (ASF, 2011), with a resolution of about 12.7 m, allows estimating the fundamental geomorphological features.The average slope of La Liboriana is 57.6 %, and the basin longitude and perimeter are 13.5 km and 57.8 km, respectively.The Strahler-Horton order of the principal stream is 5, and the longitude and slope of the principal stream are 18.1 km and 8.1 % respectively.The highest elevation of the watershed (Cerro Plateado) reaches 3609 210 meters above sea level (m.a.s.l), while the outlet of the basin is at 1316 m.a.s.l..The longitudinal profile (see Figure 3) exhibits considerable steepness in the upper basin.The 99th slope percentile of order 1 streams is 78%.For streams order 2, 3, 4, and 5 the 99th slope percentile are 61, 27, 18 y 11%, respectively.These features are typical of Andean mountainous basins.The hypsometric curve shows considerable elevation changes for a cumulative area below 33% (see blue line in Figure 3).

215
Geomorphologically, this kind of watershed is prone to the occurrence of flash floods.
At the sub-basin scale, La Liboriana exhibits a vast range of slopes and altitude differences.Fig- ure 2 shows the Height Above the Nearest Drainage -HAND-model (Rennó et al., 2008) for La Liboriana.The HAND calculates the relative difference between cell i and its nearest streamflow Vegetation and land use vary considerably within the basin.In the upper La Liboriana basin, there are crop fields and tropical forests; a portion of the watershed is considered a national park.Hill-230 slopes near the divide do not evidence significant anthropic intervention most likely due steepness of the slopes in this region.Down the hills and at the bottom of the valley there are coffee plantations (the primary economic activity of the region) and pastures.The watershed exhibits grazing areas and Table 1.Description of the soils in the region (Osorio, 2008).

Type
Slope

235
The Andean Colombian region tends to present well-developed soils of different textures with a predominance of limes and clays.One of the challenges for hydrological modeling and risk management in the country is that soils are not well mapped; national soil cartography is usually available in 1:400.000scale.At this scale, the municipality of Salgar, including La Liboriana basin, corresponds to only one soil texture.The only available information corresponds to descriptions by regional au-240 thorities.Osorio (2008) describe La Liboriana soils as well-drained with poor retention capacity.
Organic material is predominant in the first layer and clay loam soil within the second one.The depth of the soil is hillslope dependent, varying between 20cm and 1m (Osorio, 2008).Table 1 provides a summary of soil characteristics for five different categories, all as a function of slope.Each soil category has a corresponding depth and a qualitative description of permeability and retention.

245
A field campaign few days after the flash flood event was instrumental in obtaining cross-section geometry along the main channel, including at the outlet of the basin.The cross-section has a rectangular shape, and it is 4.6m wide and has a height of 5m for a total area of about 23m 2 .A visual For the technically inclined reader, the hydrological model and sub-models are written in Fortran 90, and the interface to the model, pre-process, and post-process tools are in python 2.7.The Fortran 305 code is warped to python using f2py (Peterson, 2009) and it is publicly available under the Water- shed Modelling Framework WMF in a web repository (GitHub).

Virtual Tracers
Virtual tracers are implemented into the model to discriminate streamflow source in superficial runoff have often been more qualitative than quantitative (Creutin and Borga, 2003b;Šálek et al., 2006;Borga et al., 2014).
Rainfall from radar reflectivity fields is classified into stratiform and convective using the Steiner et al. (1995) approach before the implementation of the rainfall tracer.At each time step, the model 335 takes into account the total rainfall, assuming that at each particular cell the precipitation is either entirely convective or entirely stratiform.This assumption could lead to estimation errors at basins represented by coarse cells (low DEM resolution) where convective and stratiform precipitation are likely to coexist.In the present study the spatial resolution of the DEM is 12.7m, higher than the resolution of the radar retrievals (about 125m), so the potential convective and stratiform rainfall 340 concurrence is very low, and it could not be identified using the Steiner et al. (1995) approach.Once within the model, rainfall separation follows the methodology proposed for flow type separation.

Hazards Assessment
During the 2015 La Liboriana flash flood event, two central processes took place.Shallow landslides 345 occurred in the upper basin, and a flash flood generated flooding, and human and infrastructure losses.

Shallow landslides sub-model
The shallow landslides submodel coupled to the hydrological model is proposed by Aristizábal et al..

350
The model conceptualizes the stability of a cell based on the concept of an infinite slope described by the geomorphological and soil properties of the cell.The model classifies cells into three groups: unconditionally stable, conditionally stable and unconditionally unstable.Figure 6 describes the variables of the model and the balance of forces considered.Three parameters determine stability of each cell: (i) residual soil thickness water table Z i,min estimated as in equation ( 1), (ii) the maximum soil imum slope at which the soil remains stable β i,0 estimated as in equation ( 3). (2) The model assess conditionally stable cells in function of their perched water table for each cell (Z i,w , equation ( 4)) and the critical satured depth (Z i,c ).Slope failure occurs when Z i,w is greater The model requires hydraulic parameters for all network cells to determine flash-flooding spots.

385
Channel width is estimated using the Leopold (1953) approach W i = 3.26Q i 0.469 .Channel slope (S i,0 ) is obtained as the mean value of the slopes that correspond to the cells of a hydrological reach.
The characteristic particle diameter D i,50 is assumed equal to 0.138 and constant.The cross-section (C s ) is obtained from the DEM, for every network cell, perpendicular to the flow direction of the cell (D i,8 ).

390
For each time step t and for each stream cell, equation ( 6) determines the height of the water table Y i (t) using the simulated streamflow Q i,sim (t) and flow velocity v i,sim (t).The model calculates the friction velocity (v f r,i ) using Y i as in equation ( 7) derived from Keulegan and Rouse equations (Takahashi, 1991).Equations ( 8) and ( 9) allow the estimation of the concentration (c i (t)) 395 and constitutive coefficients (r i ), respectively.The stream flow plus estimated sediments and rubble is estimated according to equation (10).
Di,50 + 6.25 9.81 0.0128 assuming infinite sediment and rubble supply, Q i,sed is the maximum stream flow for the section.
To determine the flooded area, an equivalent Y i,sed must be found in order to obtain an stream flow ( Qi,sed ) (equation ( 11)) that equals Q i,sed .The estimation of Y i,sed is an iterative process .The model obtains the flooded area (A i,sed ) based on Y i,sed and equation ( 12).

410
Qi,sed = 2 5 r i (j∆y i,sed ) Âi,sed = ∆x Resulting flood maps might evidence the presence of small isolated flood spots and discontinuities 415 at flood spots where flow direction changes from orthogonal to diagonal or vice-versa.We included two post-processing steps to correct these issues by (i) using an image processing erosion algorithm to remove the small isolated flood spots.The erosion is performed once with a 3x3 kernel, and (ii) to solve the flow direction discontinuities, each flooded cell seeks to flood its eight neighboring cells.A neighbor cell is also flooded if the altitude of the original flooded cell plus the flood depth is higher 420 than its altitude.

Hydrological runoff scheme modification
Horizontal flow equations could be either linear or potential, as shown in equation ( 13).In the modified hydrological model, β and α are estimated by the user and then set into the model.
Non-linear equations in lateral flows could result in a better representation of processes at high resolutions (Beven, 1981;Kirkby and Chorley, 1967).A non-linear approximation to runoff is presented in equation ( 14).This approximation is a modification of Manning's formula for flow in 430 gullies.According to Foster G.R. (1984), the values of ε and e 1 are 0.5 and 0.64, respectively.The non-linear equation ( 15) corresponds to an adaptation of Kubota and Sivapalan (1995) formula for subsurface runoff.
Where k s is the saturated hydraulic conductivity, and b is dependent on the soil type, and it is assumed equal to 2. A g is the equivalent cross-section area to the maximum gravitational storage (H g ).There is also return flow from tank 3 to tank 2 when S 3 = H g , representing runoff generation by saturation.

440
The equations describe the momentum of a kinematic wave approximation.In both cases, velocity depends on the tank storage.These relations could be summarized by the equation ( 13), that could be solved numerically coupled with a mass balance equation ( 16).This equation takes into account the storage at each time step (S tank (t)), the longitude of the element (∆x), the time step size (∆t), and the speed estimated for the flow in the time step (v tank (t)).Equation ( 16) is related to equation 445 (13) through the velocity term (v tank ) and the cross-sectional area (A).The solution to v tank and A is obtained through an iterative scheme.The total outflow from the tank is calculated using equation (17).
The primary results of the present study include the reconstruction of the 2015 Salgar flash flood, the assessment of the importance of soil moisture in the hydrologic response of the basin, and the 455 evaluation of the relative role of stratiform and convective precipitation cores in the generation of the observed extreme event.This section is based on the results from the analysis of the hydrological simulation, as well as shallow landslides and flash floods occurrence and simulation.training convective cores move over La Liboriana basin generating intense precipitation peaks during 2.5 hours.It is important to note that these cores did not strengthen within La Liboriana basin; these systems formed and intensified over the western hills of Farallones de Citará draining to the Department of Chocó towards Atrato river.The latter is not a minor fact because once the convective 470 system move with a northeast direction, the maximum intensity cores do not fall over the steepest hills of La Liboriana basin but rather near the basin outlet where the slopes are considerably flatter.
Figure 7b shows the spatial distribution cumulative rainfall during Event 1, corresponding to a basinaverage precipitation accumulation of 47 mm,with the maximum precipitation located towards the bottom third of the basin.Event 2, on the other hand, started as a thunderstorm training event with 475 two convective cores moving from the southeast followed by remaining stratiform precipitation.
Even though the average cumulative rainfall over the basin was 9 mm less than during Event 1, this event is characterized by orographic intensification within the basin leading to a more heterogeneous spatial distribution and with the highest cumulative precipitation in the steepest portion of the basin (see Figure 7b).Event 2 spatial distribution and highly localized observed intensities most likely led 480 to the flash-flooding episode, as it is explored in the remainder of this section.
Figure 8a presents the results of the hydrological simulation at the outlet of the basin.The model simulation is set to reach a base flow of 3 m 3 /s, a value that corresponds to the discharge measurements during field campaigns days and weeks after the flash flood event and during dry spells.

485
The simulation shows that Event 1 generates a hydrograph with a peak flow of Q max = 160 m 3 /s that did not result in casualties or infrastructure losses, but set wet conditions in the entire basin Figure 8b presents capillary storage (purple), as well as runoff (continuous blue) and gravitational (dashed blue) storage temporal variability.As expected, runoff storage is only non-zero during the storm duration, while gravitational storage increases considerably during rain events, followed by 500 a slow recession.There is an increment of basin-wide capillary storage during Event 1, remaining considerably high the time leading to the occurrence of Event 2.
In this study, we propose a graphical method to assess the soil-rainfall-discharge coupling holistically.The first step is to classify all the cells within the watershed in a predetermined number of 505 groups according to their localization and the distance to the outlet.The aim is to establish a coherent and robust spatial discretization, thus allowing to summarize the concurrent spatio-temporal variability of the different processes in 2D diagrams.Figure 9 presents an example of the spatial structure of cell grouping for La Liboriana basin using 10 (left) and 50 (right) groups.

510
Figure 10 presents the proposed 2D diagrams obtained for the simulation of the La Liboriana basin flash flood using a spatial discretization with 50 groups.Figure 10a includes the evolution of  the average rainfall over the basin (black line), and the spatio-temporal evolution of capillary storage (filled isolines) and return flow (colored isolines from white to red) by groups.For the analysis, it is relevant to highlight that higher numbered groups are located away from the outlet of the basin and 515 correspond in this case to considerably steeper slopes.Figure 10b presents the evolution of streamflow at the outlet of the basin (black line) as well as the gravitational storage (filled isolines) and runoff (colored isolines) spatio-temporal evolution.
It is well known that temporal variability of rainfall intensity plays an important role in hydro- the life cycle (Figure 11b).This change in precipitation intensity is associated with a considerable intensification of the training convective cores due to orographic effects.Events 1 and 2 also exhibit differences in the elapsed time between rainfall occurrence and streamflow increment given the rela-525 tive timing of stratiform versus convective rainfall (see the gray band in Figure 11a and b).For Event 1, the median elapsed time between rainfall and streamflow (Et p50 ) is 1.12 hours while for Event 2 Et p50 is 0.79 hours.The median elapsed time between the convective portion and the streamflow (Etc p50 ) in Event 1 is 0.75 and 0.46 in Event 2. The minimum value of the convective elapsed time Etc min also descends from 0.42 to 0.25 hours.On the other hand, there is an increase of median 530 elapsed time between stratiform rainfall and streamflow (Ets p50 ) from 1.21 to 1.83 hours.In Event 2, the convective rainfall and the runoff show a similar evolution, denoting a strong influence of the convective portion (Figure 11b).
As mentioned before, average rainfall accumulation over the basin for Events 1 and 2 is 47mm 535 and 38mm, respectively.During Event 1 (2), convective (stratiform) average accumulations are 28 (23) and 17 ( 14) mm, respectively (Figure 12a and b).The maximum rainfall intensities are relatively similar with 150mm/h and 180mm/h for Events 1 and 2, respectively.Despite this, Event 1 does not trigger a flash flood event.The overall evidence suggests that the discriminating factor between both events does not lie in the portions of convective or stratiform rainfall but rather in their 540 spatial distribution.
Figures 12a and b, and Figures 12c and d show the convective and stratiform cumulative rainfall, and the spatially-averaged convective and stratiform rainfall, both as a function of sub-basin reach, respectively.Figures 12a and b show that while Event 1 exhibit similar convective and stratiform 545 rainfall accumulation for different watershed scales, Event 2 shows a more significant cumulative contribution of convective rainfall than of stratiform precipitation.Convective rainfall tends to cover less area, and at the same time present a spatio-temporal erratic behavior (Steiner et al., 1995;Houze, 1989).Figures 12c and d provide evidence that convective rainfall present higher variations at small sub-basins than for a larger-order basin.During Event 2, convective accumulation reaches higher 550 values for small and medium sub-basins.Convective rainfall occurrence at the upper sub-basins has significant implications due to geomorphological conditions associated to zero-order sub-basins (Sidle et al., 2018).Figure 13 presents Pearson correlations coefficients between the convective and stratiform hydrograph portions and the runoff and subsurface flow.According to this, convective and stratiform rainfall exhibit a weaker relationship with the flow characteristics at small scales (under 555 5km 2 ).This is likely to be associated with increasing variability of rainfall and hydrograph formation at small scales (Ayalew et al., 2014).Correlations tend to grow with increasing area, indicating stabilization of the hydrograph formation.Additionally, subsurface flow presents higher correlations during Event 1, while correlations with runoff are higher for Event 2, highlighting the most impor- tant process in each case.The relevance of subsurface flow is likely due to the rainfall characteristics 560 during Event 1, with homogeneous rainfall intensity and high rate of basin recharge (see Figure 8b).
On the other hand, saturation processes and a wet soil profile explain the observed higher correlations during Event 2.

565
In addition to the hydrological simulation, associated hazards such as floods and landslides are also modeled and discussed in this study.The landslides model as described in the previous section  requires additional information including soil depth Z, cohesion C , friction angle φ and specific weight γ.According to the description by Osorio (2008), illustrating a clay-slime soil, C is assumed equal to 4KN , φ to 30 0 and γ to 18; Z vary with slope according to Table 1. Figure 14a presents the observed landslides triggered by Event 2 based on aerial photos and satellite images (Landsat/Copernicus, and Google) taken before and after the flash flood.Figure 14b shows, by hills, the map of total unstable cells during the simulation period, and Figure 14c shows the time series of the number of simulated unstable cells during Event 2 (continuous purple line) and 575 the mean rainfall over the basin (inverse axes, blue line).Calibration of the landslide model was performed by finding the maximum overlap between simulated and observed unstable and stable cells, and at the same time reducing the overall number of false positives and false negatives.It is important to note that the calibration strategy is not a cell-by-cell modification of the parameters involved but rather a basin-wide modification of soil properties.A sensitivity analysis of soil parameters is 580 carried out by making small variations of the variables within specified intervals: φ between 25 and 32, γ between 17 and 19, C between 3.5 and 4.2, and Z between 0.1 and 3 m .The sensitivity analysis suggests that slight variations in the parameter in Z produce significant changes in the modeled landslides, resulting in an overestimation of the number of unstable cells, or no unstable cells at all.
Following Table 1, the average soil depth in the basin is only 0.3 m, a value that corresponds to un-585 derestimation according to the inspections during field visits.For this reason, the results presented in Figure 14 use a Z map scaled by a calibration factor of 3.5, preserving the spatial dependence on the slope, but achieving a more realistic soil depth and better spatial distribution of landslide occurrence.
The model represents considerably well the spatial distribution of the areas that are prone to trig-590 ger shallow landslides during Event 2, showing a significant density of unstable cells in the hills where slides took place.However, despite the calibration efforts, the total number of unstable cells is relatively low compared to observations.A pinpoint localization of the unstable cells is still considered a hard task in part due to the small temporal and spatial scale at which landslide processes take place (Aristizábal et al., 2016;Dhakal and Sidle, 2004;Wu and Sidle, 1995).Additionally, the 595 lack of detailed soil information increases the simulation uncertainty.Notwithstanding the difficulties, the results suggest that the model simulations could have been used and should be used in the future for early detection and warning to improve both short and long-term risk reduction strategies.
According to the model results, a significant portion of the landslides happened during the first of 600 the training convective cores reaching the upper basin (see Figure 14c).Due to the high soil moisture, cells located in the mentioned region where prone to failure.The second and more intense convective core, between 1:20 and 2:15 a.m., also triggered additional slides for an extended period.
During storm events, shallow landslides become a significant source of gravels and sediments.These particles tend to reach the river network, increasing the magnitude of the streamflow and its destruc-605 tive power.Despite this, there is no current explicit link in the model between the landslides scheme and the hydraulic submodel.Nonetheless, HydroFlash assumes that sediment transport is limited by channel capacity, and not by hillslope supply, which in this case is guaranteed by the elevated The model results indicate that 11% of flood spots happen at elements of order 1 and 2, and, 18, 38 and 32% happen at orders 3, 4 and 5, respectively.This also highlights a coherent geomorphological representation of the flooded channels and hills with the order.basin response.This overall approach helps to isolate flood generating mechanisms or causative factors both in time and also in space, focusing on the important physical processes and not only on the statistics (Klemes, 1993;Merz and Blöschl, 2003).It is hoped that knowledge improvement leads to the anticipation of the warning and response by risk management entities.

645
The evolution of the simulation of Events 1 and 2 show evidence of remarkable behavioral differences.During Event 1 both gravitational and capillary tanks are filled along and across the basin as a result of the quasi-homogeneous rainfall spatial distribution.The return flow is low, and most of the runoff occurs within the first 20 groups (40% of the watershed closest to the outlet).In the period between both events, there is a recession in the capillary and gravitational storages in the entire  2010), low infiltration increase runoff rates, which finally affects the susceptibility of the basin to flash floods occurrence (Wagner et al., 1999;Penna et al., 2011;Tramblay et al., 2012).La Liboriana geomorphological characteristics, corresponding to a steep tropical basin, determine the potential energy that controls water transit velocity.Water tends to reach faster the channels in order 1 and 2 hills, and, at the same time, the sediment production and transport in these hills tend to be 680 larger.Order 3 sub-basins most likely act as transport elements, with no important energy losses.
Floods tend to occur in order 4 and 5 sub-basins due to the widening of the channel and slope atten-  (2016) states that in addition to the rainfall, geomorphological characteristics and antecedent soil conditions are key in the generation of flash flooding.

695
However useful, the evidence in this work only takes into account two successive events; an analysis of more cases and different spatial scales (different basins) would provide robust conclusions in this direction.It is clear that it is not conclusive enough to focus on a single extreme event, rather than on a spectrum of floods Merz and Blöschl (2003).The results suggest it is imperative to study in depth the long-term link between the relative basin and drainage network orientation and the pre-700 ferred path of precipitation events and its role in defining the frequency of flash flood occurrence.
A better understanding of the network-hills-preferential rainfall advection structure could provide information about basins prone to flash floods when information is scarce.functional early warning systems.There is a global need for better knowledge and understanding of the hydrological and meteorological conditions that, combined, lead to the manifestation of natural hazards.Such understanding must result in useful practical applications that improve risk management practices saving lives.
In the case of La Liboriana flash flood, radar reflectivity fields were available from a C-Band radar operated by the Early Warning System of Medellín and its metropolitan area as part of a local risk management strategy.While the municipality of Salgar is located far outside Medellín's metropolitan area, the radar is about 60 km away from Salgar, and the reflectivity retrievals enable the classification of precipitation fields into convective and stratiform areas using widely accepted methodologies by the meteorological community.Radar reflectivity is also a proxy for precipitation allowing a quantitative estimation.This estimation was used together with the hydrological modeling to assess the different basin-wide processes taking place during the flash flood triggering rainfall event.
The overall methodology reproduces considerably well the magnitude and timing of La Liboriana flash flood discharge peak, as well as the areas of regional land-slide occurrence and flood spots location.Simulation results indicate that the flash-flood and the regional land-slide features were strongly influenced by the antecedent rainfall associated with a northeasterly stratiform event that recharged the gravitational and capillary storages.The hydrologic simulation shows that the antecedent event set wet conditions in the entire basin before the occurrence of the flash flood event, governing the streamflow during the latter.The first of the two successive convective cores (thunderstorm training) over the same region during the flash flood event saturated both capillary and gravitational storages in the upper part of the basin and generated both return flow and significant runoff.The second convective core resulted mainly in surface runoff spatially collocated with the steeper slopes, generating the kinetic energy needed to produce La Liboriana flash flood.Overall results also show an excellent agreement between the simulated flood spots and the observed ones; this in spite of the limitations imposed by the resolution of the DEM used for extracting cross sections, and the model simplifications described in the Methodology section.
The described model-based approach is useful to isolate flood generating mechanisms and as a tool for policy-makers not only for short-term decisions in the context of an early warning system but also as a planning resource for long-term risk management.While several improvements could be implemented, including a better representation of hydraulic parameters, and a link between landslides and flood spots similar to the strategy in the STEP-TRAMM model (Fan et al., 2017), the results suggest it is possible to use the low-cost methodology introduced here to not only improve our understanding of hydrological processes but also as a risk management tool in countries and re-

Between
May 15th and May 18th, 2015, several storms took place over La Liboriana basin located in the municipality of Salgar, in the department of Antioquia (Colombia).By 2015, the population of Salgar was estimated at 17400 persons, 8800 residing in the urban area.La Liboriana is a small tropical basin (56km 2 ) located in the western range (Cordillera Occidental) of the Colombian Andes.During the night of May 17th, between 02:00 and 09:00 A.M. (local time), a precipitation event 135

Figure 1 .
Figure 1.Example of infrastructure damage as a result of the La Liboriana May 18 of 2015 flash flood event.a) Aerial photograph before the event (2012) taken during a mission of the Department of Antioquia´s Government, and b) satellite image after the event (2015-05).The images show the destruction of most houses in that particular community, a bridge over La Liboriana, and the main road.The images also present changes in the delineation of the main channel as well as considerable erosion in the river margins.

190
the overall skill of the proposed methodology.The document is structured as follows.Section 2 describes in more detail the region of study, La Liboriana basin, including geomorphological and climatological characteristics of the basin as well as the information sources used in this study.Section 3 presents a description of the overall 195 methodology and the model used for the reconstruction of the 2015 La Liboriana flash flood event, including flow separation, shallow landslide parameterization, and the proposed hydraulic model.Section 4 describes the main results of the study, including model parameterization and validation.

Figure 2 .
Figure 2. Geographical context of Liboriana basin, located in Colombia, in the Department of Antioquia.The color bar represents the Height Above the Nearest Drainage (HAND).Low HAND values correspond to areas prone to flooding.
220 cell j.La Liboriana HAND exhibits values between 500 and 800m.Near the outlet of the basin, over the banks, there are values close to 0 m.High HAND values at the upper region of the watershed often denote areas of high potential energy, with increased sediment production and frequent shallow landslides occurrence.Banks with low HAND values are more susceptible to flooding and tend to correspond to areas prone to extensive damages caused by extreme events.While the eleva-225 tion differences described in Figure 2 are typical of the region, the social challenges lie in the high vulnerability of the Salgar given the location of the main urban settlement.

Figure 3 .
Figure 3. Longitudinal profile of the principal stream (colored line) from the upper part of the basin to its outlet at the entrance of the urban area, hypsometric curve of the drainage basin (blue line).

Figure 4 .
Figure 4. Channel cross-section shown an example of flooded infrastructure after the flash flood event.The section shows mud marks on the walls with heights varying between 0.5 and 1.2 m.These mud stains are evident in buildings located 4-5 m away from the channel.The photograph also shows the width of the channel and the total estimated depth during the flash flood.

Figure 5 .
Figure 5. Illustrative diagram of the methodology followed in the present study.The left column represents the availability of radar-based QPE and a detailed DEM as the basis of the modeling framework and the mains aspects of the distributed hydrological model used.In each cell five tanks represent the hydrological processes including capillary (tank 1), gravitational (tank 2), runoff (tank 3), baseflow (tank 4) and channel storage (tank 5).The state of each tank varies in function of vertical and lateral flows as shown in the diagram, where the storage is represented by Si, the vertical input by Di which in turns depends on Ri.Ei represents the downstream connection between cells, and evaporation.The implementation of convective and stratiform rainfall separation and virtual tracers are presented in the top two right panels.The implementation of the shallow landslides model and HydroFlash are schematized in the bottom two right panels.
310and subsurface flow and to assess the portion of streamflow from convective rainfall and stratiform precipitation, recording at each time stem and each cell the source of water.The model archives the results of the virtual tracing algorithm at the outlet of the basin and at each reach allowing the study of the role of flows from different nature during extreme events at different spatial scales to get more insights about the soil-driven flow regulation.The soil texture and composition could play diverging 315 roles during extreme events.The soil could work as a porous medium in which the water travels fast, making the subsurface flow an essential part of the hydrograph, or on the other hand, its saturation could lead to a decrease in infiltration rates and a potential increase of surface runoff generation.The flow separation module operates in tanks 2 (runoff storage) and 3 (subsurface storage).The 320 module marks water once it reaches any of those two tanks and the runoff-subsurface flow percent-Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-452Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 25 September 2018 c Author(s) 2018.CC BY 4.0 License.age are taken into account once the water enters tank 5 (the channel).At this point, the scheme assumes that the water in the channel is well mixed, implying that the flow percentage is constant until a new inflow enters the channel.325 Both spatial and temporal rainfall variability influence the hydrograph Shope (2016).Stratiform rain cells cover large areas and are related to long duration hydrographs with low peak flow.On the other hand, convective rainfall cells cover limited areas and are related to short duration hydrographs with high peak flows.Stratiform rainfall tends to recharge the basin, while convective tends to increase runoff generation and the formation of the hydrograph.In the literature, these associations 330

Figure 6 .
Figure 6.The geotechnical conceptual model.γ is the soil bulk density, γw is the water density, Zw is the saturated soil thickness above the slip surface, Z is the soil thickness measured vertically, β is the gradient of the hillslope.And QL and QR are the resultant forces on the sides of the slice.Figure and description are adapted from Aristizábal et al. (2016).
370 where S 3,i represents gravitational storage, W s and W f c the soil saturation and field capacity respectively, γ represents soil density, φ is the soil stability angle, C the soil cohesion and β the cell slope.14 Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-452Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 25 September 2018 c Author(s) 2018.CC BY 4.0 License.Flash flood submodel (HydroFlash)It is common to use complex hydraulic 2D and 3D models, such as Iber(Cea and Blade, 2015) or 375 HEC(Gibson et al., 2010) for flood spots identification.This kind of modeling requires detailed information about the terrain and river bathymetry and are usually implemented for a particular river cross section due to the high information demand and high computational costs, precluding the operational implementation of 2D and 3D methodologies in regions with scarce resources.Also, this traditional hydraulic methodology tends to analyze only some network elements, overlooking the 380 complexity of the basin.In this work, we introduce a low-cost 1D hydraulic model for flash flood simulation referred to as HydroFlash.The model extracts the cross-profile from the DEM for each cell considered part of the network, estimating flood spots at each network cell during execution time.

Figure
Figure 7a presents the temporal evolution of the convective-stratiform rainfall partitioning during

Figure 7 .
Figure 7. a) Temporal evolution of the convective-stratiform rainfall partitioning during both Events 1 and 2. The figure shows the total rainfall (yellow), and the convective (blue) and stratiform (green) portions integrated over La Liboriana basin.b) and c) Spatial distribution of the cumulative rainfall during Events 1 and 2 over La Liboriana basin.

Figure 8 .
Figure 8. Summary of the Results from the hydrological simulation.a) Simulated streamflow, runoff and subsuperficial flow separation and convective-stratiform generated discharge discrimination.b) Mean runoff, gravitational, and capillary storages during the simulation period.

Figure 9 .
Figure 9. Example of watershed grouping as a function of distance to the outlet for La Liboriana basin.The left panel corresponds to clustering using ten different groups, and the right panel shows the 50-groups categorization used in the present study.

Figure 10 .
Figure 10.Spatiotemporal analysis of Events 1 and 2. a) Simulated capillary moisture (filled green-to-blue contours) and returned flow occurrence (white to red isolines).The black line represents the average rainfall over the basin.b) Simulated gravitational moisture (filled green-to-blue contours) and runoff (yellow-to-red isolines).The black line represents streamflow at the outlet of the basin.The green-to-blue color bar serves as a reference for capillary moisture and gravitational water content.

520
graph structure.During Event 1 rainfall accumulated over the basin at a relatively stable rate (Figure11a).On the other hand, Event 2 presents a significant increase in rainfall rate in the second half of 20 Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-452Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 25 September 2018 c Author(s) 2018.CC BY 4.0 License.

Figure 11 .
Figure 11.Accumulated rainfall and streamflow for a) Event 1 and b) Event 2. The accumulation is expressed in percentage respect to the total value in each case.Median elapsed time and minimum elapsed time () are estimated between total (Etp50, Etmin), convective (Etcp50, Etcmin), and stratiform (Etsp50, Etsmin) rainfall and the runoff portion of the streamflow.Gray bands correspond to the periods for elapsed time estimation.

Figure 12 .
Figure 12.Cumulative rainfall versus the area at each reach of the basin for a) Event 1 and b) 2. Panels c) and d) show the average rainfall versus the area at each reach of the basin.Yellow dots correspond to the total, green dots to convective and blue dots to stratiform rainfall.All panels use a logarithmic scale for the basin area.

Figure 13 .
Figure 13.Pearson correlations among convective and stratiform portions of the rainfall and runoff and subsurface flow for different reach areas.Dashed lines correspond to Event 1 and continuous lines Event 2.

Figure 14 .
Figure 14.a) Observed landslides triggered by Events 1 and 2. The figure is based on aerial photos and satellite images (Landsat/Copernicus, and images available on Google) taken before and after the flash flood event.b) Map of total unstable cells during the simulation period.c) Time series of number of simulated unstable cells during Event 2 (continuous purple line) and mean rainfall over the basin (inverse axes, blue line).

Figure 15 .
Figure 15.Simulated flood spot at the peak of Event 2 in different locations.a) Basin drainage network.White squares correspond to regions of interest highlighted in panes b) to f). b) Zoom at the outlet of the basin, where an important portion of the human and infrastructure loses took place.c) Zoom at La Margarita settlement also affected by the flash flood.d) to f) Zoom at key locations along the principal stream.Observed flood spots are shown in blue polygons and model flood spots in red to white grids.

650basin.
Capillary storage decays considerably slower than gravitational storage.During Event 2, the flash flood event, the first convective core saturates both capillary and gravitational storages in the upper part of the basin and generates both return flow and significant runoff.Due to soil saturation, the second convective core results mainly in surface runoff.During this event, runoff is generated the lower part of the basin while extreme runoff rates are evident in the upper part of the basin, col-655 located with the steeper slopes.On the other hand, subsurface flow is more important in magnitude than runoff describing Event 1, while runoff is more relevant for Event 2. The precedent storage and the presence of thunderstorm training profoundly condition the streamflow during Event 2. The overall evidence suggests that precedent capillary moisture in the basin plays an essential role in modulating river discharge.This behavior could be linked to the temporal occurrence and relative 660 importance and timing of stratiform and convective formations previously described.While convective and stratiform partitioning could influence the runoff and subsurface flow separation, the spatial distribution of rainfall relative to watershed network morphometry structure impose a condition on the hydrological response of the basin.In other words, hydrograph formation is 665 not only determined by the rainfall accumulation or maximum intensity, but also by its spatial structure.The structure of the rainfall associated with La Liboriana event highlights the need to consider in more detail the role of orographic rainfall intensification in practical applications such as early warning systems.Evidence suggests the spatial structure of the rainfall is at least as important as the geomorphological features of the basin regulating the generation of flash flood events.670 An integrated spatiotemporal analysis of rainfall evolution, together with soil storages in the basin is necessary to study the relevance of antecedent conditions and precipitation type, intensity, and 27 Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-452Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 25 September 2018 c Author(s) 2018.CC BY 4.0 License.location in the generation of flash flood events.Event 1 increased the overall soil moisture with an associated decrease on infiltration rates similar to results reported by Penna et al. (2011) and Zehe 675 et al. ( focused on trying to understand the general causative factors behind the 685 occurrence of flash floods finding similar to our results, a significant role of basing geomorphology, orography and local convection.For example,Lehmann and Or (2012), using a shallow landslide model, finds an important role of the topography and the rainfall conditions.Turkington et al. (2014) shows how intense locally driven convection appears to be the main meteorological trigger for flash occurrence in the French Alpes.Camarasa-Belmonte (2016) shows how rainfall intensity and dura-690 tion influences the shape of the hydrograph, with intense rainfall shortening the response time of the basin, and large durations increasing the flood peak.In the Mediterranean region,Boudou et al.

705
Extreme rainfall events such as the one that triggered La Liboriana tragedy frequently take place in Colombia and the entire global tropical belt over ungauged basins, often triggering flash floods and torrential flows, endangering vulnerable communities due to poor long-term planning and lack of Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-452Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 25 September 2018 c Author(s) 2018.CC BY 4.0 License.
Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-452Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 25 September 2018 c Author(s) 2018.CC BY 4.0 License.gions with scarce resources.In the case of Salgar, there was radar information available, but in many cases, the only near-real-time precipitation information comes from satellite retrievals.It would be very useful from a practical point of view to assess in which cases satellite information is enough for risk management applications.Comparing the Salgar case with the Mocoa (Department of Putumayo, Colombia) April 1st, 2017 flash flood that resulted in more than 330 fatalities could help 750 address this question.7 Acknowledgements This work was supported by SIATA (Sistema de Alerta Temprana de Medellín y el Valle de Aburrá) funds provided by Area Metropolitana de Medellín y del Valle de Aburrá (AMVA), Municipio de 755 Medellín, Grupo EPM, and ISAGEN under the Research and Technology Contract CD511 of 2017.Nicolás Velásquez was partly funded by Universidad Nacional de Colombia under the Facultad de Minas graduate scholarship program.30 Hydrol.Earth Syst.Sci.Discuss., https://doi.org/10.5194/hess-2018-452Manuscript under review for journal Hydrol.Earth Syst.Sci. Discussion started: 25 September 2018 c Author(s) 2018.CC BY 4.0 License.