Hydrometeorological Conditions Leading to the 2015 Salgar Flash Flood: Lessons for Vulnerable Regions in Tropical Complex Terrain

Flash floods are a recurrent hazard for many developing Latin American regions due to their complex mountainous terrain and the rainfall characteristics in the Tropics. These regions often lack the timely and high-quality information needed to assess, in real-time, the threats to the vulnerable communities due to extreme hydrometeorological events. The systematic assessment of past extreme events allows improving our prediction capabilities of flash floods. In May 2015, a flash flood in La Liboriana basin, municipality of Salgar, Colombia, caused more than 100 casualties and significant infrastructure 5 damage. Despite the data scarcity, the climatological aspects, meteorological conditions, and first-order hydrometeorological mechanisms associated with La Liboriana flash flood, including orographic intensification and the spatial distribution of the rainfall intensity relative to the basin’s geomorphological features, are studied using precipitation information obtained using a weather radar quantitative precipitation estimation (QPE) technique, as well as from satellite products, in situ rain gauges from neighboring basins, quantitative precipitation forecasts (QPFs) from an operational Weather Research and Forecasting 10 (WRF) model application, and data from reanalysis products. La Liboriana flash flood took place during a period with negative precipitation anomalies over most of the country as a result of an El Niño event. However, during May 2015, moist easterly flow towards the upper part of La Liboriana caused significant orographic rainfall enhancement. The overall evidence shows an important role of successive precipitation events in a relatively short period, and of orography, in the spatial distribution of rainfall and its intensification as convective cores approached the steepest topography. There were three consecutive events 15 generating significant rainfall within La Liboriana basin, and no single precipitation event was exceptionally large to generate the flash flood, but rather the combined role of precedent rainfall, and extreme hourly precipitation triggered the event. The results point to key lessons for improving local risk reduction strategies in vulnerable regions with complex terrain.

country due to the fact that important human settlements frequently occupy floodplains.
Flash floods are associated with short-lived, very intense convective precipitation events, usually enhanced by the orography, 35 over highly saturated land surfaces with steep terrains (Šálek et al., 2006;Llasat et al., 2016;Douinot et al., 2016;Velásquez et al., 2018). Flash floods are highly destructive, often resulting in significant human and economic losses, making them one of the most catastrophic natural hazards (Jonkman, 2005;Roux et al., 2011;Gruntfest and Handmer, 2001). Jonkman (2005), based on information from the International Disaster Database, shows that between 1975 and 2001 a total of 1816 worldwide freshwater flood events killed over 175 thousand people and affected more than 2.2 billion people. These events not only caused 40 human and economic losses but also damages to ecosystems and loss of historical and cultural values. In Colombia, there have been several flash flood events in the last decade associated with large-scale climate forcing patterns and with isolated extreme precipitation events. The 2010-2011 La Niña event triggered more than 1200 flooding events, affecting the lives of more than 3 million people (around 7% of the country's population) and causing damages estimated in more than 6.5 billion US dollars (UN-CEPAL, 2012). 45 One of the most critical challenges associated with flash floods is their simulation and prediction with useful lead times (Yamanaka and Ma, 2017;Borga et al., 2011;Marra et al., 2017;Hardy et al., 2016;Ruiz-Villanueva et al., 2013). Skillful forecasts of the likelihood of the occurrence of a flash flood require a deeper insight into their triggering processes (Klemeš, Mora et al., 2008;Bookhagen and Strecker, 2008).
In addition to being the primary flash flood triggering factor under current climate conditions, potential changes in extreme precipitation frequency under climate change scenarios could plausibly increase flash flood recurrence around the globe (Hapuarachchi et al., 2011;Field et al., 2012), and in particular in the Tropics, although the confidence in the projection of the magnitude of the precipitation changes is low (IPCC, 2014). Local in situ evidence suggests that, while there are no long-term 70 trends in the yearly cumulative precipitation in the Department of Antioquia, short-lived events show long-term increments in intensity and frequency (Urán, 2016;Urán et al., 2019), with a substantial reduction of the return period of extreme events with implications for engineering and risk management.
As it is the case of La Liboriana extreme event, in many regions in the Tropics and around the developing world, watersheds prone to flash floods are usually located in rural mountainous areas, with scarce or non-existent real-time hydrometeorological 75 information, imposing a challenge for their prediction, modeling and, consequently, optimal risk management (Marra et al., 2017). The use of quantitative precipitation estimation (QPE) tools based on ground-based weather radar and satellite-based information for flash flood applications could potentially offset the lack of in situ precipitation products in small poorly gauged basins, becoming an important tool for the improvement of the state-of-the-art understanding of flash flood-related processes such as orographical enhancement of extreme rainfall and runoff generation (Creutin and Borga, 2003;Wagener et al., 2007). 80 Additionally, radar information is also important since antecedent rainfall serves as a surrogate of soil moisture, and different authors have shown that antecedent soil moisture significantly modulates the occurrence of flash floods (Tramblay et al., 2012;Rodriguez-Blanco et al., 2012;Coustau et al., 2012;Wagener et al., 2007;Castillo et al., 2003). Notwithstanding the limitations of radar retrievals (Šálek et al., 2006;Hardy et al., 2016), modern approaches combining radar information, in situ precipitation data, and model simulations are promising (Braud et al., 2016). Velásquez et al. (2018) present a hydrological 85 modeling framework for the reconstruction of La Liboriana flash flood assessing the runoff generation processes, concluding that the flash flood and the associated regional land-slides in the region were strongly influenced by the observed antecedent rainfall recharging the gravitational and capillary storages in the entire basin.
The aim of this study is to document the climatological aspects, meteorological conditions, and first-order hydrometeorological mechanisms triggering the 2015 La Liboriana flash flood, including orographic intensification and the spatial distribution 90 of the rainfall intensity relative to the basin geomorphological features. We focus on the aspects related to the recurrence of the event and highlight the key lessons that should be incorporated into local risk reduction strategies for other vulnerable regions with similar climate features and terrain complexity. In spite of the data scarcity, the systematic analysis of the observational evidence of the successive rainfall events triggering La Liboriana flash flood, and the output of limited-area numerical prediction models, provide an interesting case of study to improve our understanding of the main hydrometeorological factors modulating the occurrence of these events and their likelihood of occurrence. This type of study is useful in the context of policy-making, not only for short-term early warnings but also as a planning resource for long-term risk management and resilience building strategies.
The present work is structured as follows. Section 2 describes the region of study, La Liboriana basin, as well as the information sources used in this analysis. The assessment of the overall climatological and meteorological conditions, and hy-100 drometeorological mechanisms triggering La Liboriana flash flood is presented in section 3 using precipitation information derived from a weather radar QPE technique, as well as from satellite products, in situ rain gauges from neighboring basins, quantitative precipitation forecasts (QPFs) from an operational Weather Research and Forecasting (WRF) model application, and data from reanalysis products. Finally, the discussion and most important conclusions are presented in section 4.
2 Study region, data, and methods 105

Geographical context
The municipality of Salgar is located in the southwest of the Department of Antioquia, in the westernmost branch (Cordillera Occidental) of the Colombian Andes. Figure 1a

In situ rain gauges
Before the occurrence of the flooding event, there were no rain gauges available in La Liboriana basin, nor in the municipality of Salgar. Figure 1b shows the location of four rain gauges from the Colombian National Weather Service (IDEAM) in nearby regions to La Liboriana basin. The records from in situ gauges are available with daily resolution and are useful to characterize the rainfall during May 2015 in a climate context. Figure 1b also shows the location of two rain gauges installed after the 120 flooding event by the Sistema de Alerta Temprana de Medellín y el Valle de Aburrá (SIATA), the local early warning system of the Department of Antioquia's capital, with the purpose of validating the rainfall estimates using QPE techniques.

C-Band radar QPE
In the absence of in situ rainfall records during La Liboriana flash flood, we use precipitation estimates based on an empirical QPE technique described in Sepúlveda (2016) and Sepúlveda and Hoyos Ortiz (2017)   C-band polarimetric and Doppler weather radar manufactured by Enterprise Electronics Corporation. The method uses radar reflectivity retrievals and in situ disdrometer and rain gauge information to finally obtain precipitation. The method is multistage, and it is based on (i) a regression of the atmospheric volume radar reflectivity into in situ disdrometer reflectivity, and (ii) a regression between in situ reflectivity and rainfall intensity. Figure 1a (zoom 1) shows a 120 km-radius area from the weather radar (C-Band) installation site. The radar scanning strategy, which includes four plan position indicator sweeps (PPIs) at 130 0.5º, 1.0º, 2.0º, and 4.0º, and four range height indicator sweeps (RHIs), allows estimating precipitation information every five minutes with a spatial resolution of about 128 m using the 1.0º PPI. The uncertainty associated with the QPE technique is relatively low in a 120 km radius from the installation site. The QPE technique was validated for La Liboriana basin using hourly and daily information from the two in situ gauges installed after the flooding event. Figure 2 shows a correspondence between the hourly and daily cumulative precipitation estimated using the QPE technique and the precipitation registered in 135 situ. The correlation between the hourly and daily rain gauge records and the QPE estimations are, respectively, 0.65 and 0.74, indicating high reliability of the derived radar precipitation.

Convective and stratiform precipitation
Radar reflectivity fields are also used to describe the spatio-temporal evolution of the precipitation events leading to the La Liboriana flash flood and to assess the partition into its convective and stratiform portions, using the classification methodology 140 proposed by Houze (1997), andSteiner et al. (1995), which is based on the intensity and sharpness of the reflectivity peaks. Lightning activity is also analyzed together with radar reflectivity fields using data from in situ instruments that are part of a total lightning detection system developed at the University of Munich (LINET -Lightning NETwork) that locates cloud discharges and cloud-to-ground strokes (Betz et al., 2008(Betz et al., , 2009. The efficiency in the detection of the two types of strokes is greater than 90%, identifying the time of occurrence with an accuracy of 100 ns.

Predominant propagation of precipitation events
In a complex-terrain setting, the orientation of the basin relative to the wind direction and to the predominant advection of the precipitation systems determines, to a significant extent, the potential for orographic enhancement. In order to evaluate this potential, the Li et al. (1995) methodology, based on the work by Tuttle and Foote (1990), is used to assess the prevailing direction in which precipitation systems approach La Liboriana basin. We estimated the velocity fields from precipitation 150 retrievals (QPE technique) using a pattern cross-correlation technique and a variational approach, to satisfy continuity, for all the precipitation events from May 2014 to August 2018. In this case, a precipitation event is defined when the average rainfall intensity over La Liboriana basin is higher than 0.2 mm h −1 , given that this threshold was not reached in the previous hour.

Satellite-based precipitation products
In the case of La Liboriana flood, there is radar information available; however, for several basins around the world, the only 155 spatially distributed precipitation information available in near real-time is from satellite retrievals. From a risk management point of view, it is beneficial to assess how useful the satellite information is to: (i) represent the spatio-temporal structure of the precipitation events triggering flash floods, and (ii) to issue advanced warnings to the public. For this assessment, we use the 3-hourly Tropical Rainfall Measuring Mission (TRMM) 3B42 v7 product, which is in good agreement with in situ stations globally and regionally (TRMM, 2011;Kummerow et al., 1998;Huffman et al., 2007;Ceccherini et al., 2015), and 160 the half-hourly Global Precipitation Mission IMERG v05 Final Precipitation (Huffman, 2017;Huffman et al., 2007;Joyce and Xie, 2011). The TRMM-3B42 v7 product has a 0.25 • by 0.25 • spatial resolution, and a spatial coverage spanning the entire zonal band from 50 • S to 50 • N. The IMERG v05 Final Precipitation has global coverage and a spatial resolution of 0.1 • by 0.1 • . We also use GOES-13 4km spatial resolution infrared brightness temperatures (UCAR/NCAR-EOL, 2015) as a proxy for cold cloud tops, also useful to identify locations where precipitation is likely occurring.

Sea surface temperature and reanalysis products
Monthly sea surface temperature (SST) information from the NOAA Optimum Interpolation SST V2 (Reynolds et al., 2002), with 1 • horizontal resolution, is used to describe the climate context in which La Liboriana flood occurred and the long-term link between local precipitation and global SSTs. Similarly, atmospheric information (meridional and zonal winds, geopotential height and specific humidity) at different pressure levels is obtained from the ERA5 global reanalysis project with a spatial 170 resolution of 30km (ECMWF, 2017).

Back trajectory analysis
Precipitation is, among others, a function of the net moisture influx into a region of interest. In the Tropics, the divergence/convergence of moisture fluxes and the migration of the ITCZ control the net moisture influx in a control volume.
In long-(sustainable and resiliency-based urban/rural planning) and short-term (early warning) risk management applications, 175 it is remarkably important to identify the moisture sources in a region to fully characterize the local hydrological cycle and

Cumulus
Tiedke scheme the present and future potential hazards linked to extreme events. A Lagrangian approximation, such as the estimation of back trajectories, is an optimal tool to identify the moisture sources for a region (Gimeno et al., 2012), allowing to characterize its climatology and to identify spatio-temporal patterns associated with the essential moisture sources in different time-scales (Dirmeyer and Brubaker, 2007;Drumond et al., 2008;Viste and Sorteberg, 2013;Drumond et al., 2014;Huang and Cui, 2015; 180 Ciric et al., 2016;Stojanovic et al., 2017;Hoyos et al., 2018). The back trajectories, together with the specific humidity, are estimated using the three-dimensional wind and specific humidity fields from ERA5 global reanalysis, and are a proxy for the origin of the moisture generating the observed precipitation events.

WRF Model
The WRF Model, version 3.7.1 (Skamarock et al., 2008) is the foundation for the operational numerical forecasts issued by 185 SIATA on a daily basis. The model configuration uses three nested domains with 18 (191 x 191), 6 (82 x 118) and 2 (136 x 136) km horizontal resolution, and 40 vertical levels up to 50-hPa. The first domain (18 km), from 10ºS to 20ºN and from 60ºW to 90ºW, covers the entire geography of Colombia, the Caribbean Sea, the Colombian sector of the Pacific Ocean, and the Amazonia, in order to include the main external forcing factors of atmospheric circulation and precipitation over the territory.
The second domain (6 km) includes the Andean region of Colombia (1º to 10ºN, and 72º to 78ºW). The third and last domain

190
(2 km) is centered around the Aburrá Valley and extends from 5º to 7.5ºN, and 74.5º to 76.8ºW. The municipality of Salgar is located within the third domain. The model runs use the output from the 12UTC 0.5º Global Forecast System (GFS) as initial and boundary conditions. The integration time step is 90, 60, and 10 s in the 18, 6, and 2 km resolution domains, respectively.   gauges is shown in Figure 3. Most of the rainfall (more than 85%) took place after May 12th, with the heaviest period between May 12th and May 22th. The highest daily precipitation recorded by the rain gauges corresponds to accumulations around 200 55-60 mm. These daily values do not appear to be unusually high, corresponding to daily precipitation percentiles between 75th and 90th. This suggests that the available in situ records do not capture the unique spatio-temporal rainfall features that triggered the flash flood. While in situ rain gauges are perhaps the best method to measure precipitation intensity and total volume, their generally sparse nature limit their use in risk management applications, pointing to the need to combine information of local nature, or ground truth, with tools that provide spatio-temporal precipitation estimates such as satellites 205 and weather radars.
Similarly, a climatological assessment suggests that the monthly cumulative precipitation during May 2015, rather than being anomalously high as a result of regional or global-scale forcing, it corresponds to magnitudes close to the May climatology in the region. Figure (Poveda et al., 2005(Poveda et al., , 2006. According to in situ historical records, the annual precipitation in the region around La Liboriana is between 2400 and 3050 mm, and during the peak months, precipitation reaches between 200 and 350 mm.
Precipitation data from TRMM shows that during May, the mean cumulative precipitation is 415 mm with a standard deviation 215 of 170 mm. La Liboriana flood occured during May, month climatologically corresponding to the highest precipitation in the region. The QPE rainfall over La Liboriana basin appears to be higher than in nearby regions, and higher than the TRMM precipitation record. Although the period is not the same, these differences are linked to the orographic rainfall enhancement in the upper part of the basin as it will be discussed in the following subsections.

c) Precipitation anomalies over
La Liboriana, after filtering all variability with periods equal to or shorter than 13 months. d) 10-year windowed moving correlation between the Multivariate ENSO Index (MEI) and in situ precipitation. e) Spatial distribution of the correlation between filtered SST and in situ precipitation. f) Regional cumulative precipitation from the TRMM 3B42 product (blue-to-red shading) and mean SST (green-to-white shading and black lines) during May 2015. g) Cumulative precipitation (red-to-white shading) and mean SST (green-to-pruple shading and black lines) anomalies relative to the long-term May conditions. Temperature contours are every 0.5ºC timated; however, it is important to note that in situ rainfall stations are not located in La Liboriana basin. We assess TRMM precipitation considering that in many vulnerable places, especially in developing countries, it is the only information available in near-real-time. Figure 4b shows the scatter plot between monthly TRMM records and one of the in situ gauges, and the Pearson correlation between both series. The correlation between TRMM and the other three in situ gauges is similar in magnitude as the one shown in Figure 4b. The evidence suggests that, while TRMM appears to overestimate the monthly rainfall, the  Figure 4f shows the regional cumulative precipitation from the TRMM 3B42 product and mean SST during May 2015, and Figure 4g shows their anomalies relative to the long-term May conditions. The mean pattern shows warm ocean surface water north of the Equator, south of Central America between 80 and 110 W, and important precipitation over the ocean, also north of the equator, but not necessarily collocated with the 240 warmer surface waters, but rather with the 37-39º C isotherms since global circulation plays a key role in determining the overall location of the convection hotspots (Hoyos and Webster, 2012). Over the ocean, the maximum precipitation is located around 10N and 110W. There is also precipitation over Colombia, with larger values over the Amazon basin and, in particular, over the Colombian Pacific region. However, precipitation anomalies show that during May 2015, most of the country had lower precipitation than the long-term average.

245
It is important to note that, while the climate conditions, in general, modulate the likelihood of extreme event occurrence (Haylock et al., 2006;Orlowsky and Seneviratne, 2012), La Liboriana flash flood happened during El Niño conditions and the resulting negative precipitation anomalies over Colombia. From a risk assessment point of view, even under external forcing leading to less precipitation than the expected value, the likelihood of hazard occurrence is never zero: it is not sufficient to consider the monthly cumulative precipitation, but it is necessary to study the nature of the precipitation events adding up to     Figure 8f shows an area of relatively low values of Tb (e.g., Tb<200 K) associated with a broad deep cloud system located in the northeast region of the Department of Antioquia, and a small cloud system in the southwest that is likely related to the storms that occurred in the day of the disaster in the town of Salgar. The small footprint of these storms indicate the precipitating elements that produced the observed rainfall correspond to early stages of localized intense convective 305 storms. Later (not shown here), GOES imagery shows growth in the area of the cloud system, as the hydrometeors inside the cloud were lofted to the upper troposphere likely producing significant ice concentration, reducing the Tb measurements.
Twenty four hours before the flash flood (Figures 8c and d), the precipitation event was associated with a low-pressure center near the Pacific coast of Colombia, showing convergence in the region over the western slopes of the Andes in the Pacific region of Colombia, affecting indirectly the municipality of Salgar. This feature is a characteristic pattern of the synoptic conditions 310 associated with the occurrence of broad MCSs over this region Zuluaga and Houze, 2015). In fact, the cloud system in Figure 8d displays a considerable growth over the western region of Colombia, favoring humidity convergence and the generation of storms over the Liboriana basin and the town of Salgar. In addition to the May 18 and 17 events, there was a prior important precipitation event during May 15 synoptically similar to the event during the 18th, but more intense (Figures   8a and b). While the overall geopotential height was lower in the entire region, there was a marked low-pressure system that was 315 located over the Pacific coast of Colombia, the Department of Antioquia and Panama, favoring the southeasterly flow towards the western slopes of the Andes Cordillera, and even cyclonic anomalies of relative vorticity over Salgar. These conditions generated the growth of extensive cloud systems over the region with significant stratiform and convective rainfall. Figure 9 shows the cumulative precipitation during the duration of the three events in Figure 8, obtained using the radar QPE Liboriana was less than 80 mm, well below climatological values. Regardless of its magnitude, the precipitation during period 1 did not cause flash flooding in Salgar, partly because of its long-duration and, in average, relatively low-intensities (despite having high intensity spells), but more importantly because the gravitational storage in the basin's soils was low due to the below average rainfall between the first and the 13th of May (Velásquez et al., 2018). Overall, the Figure 10 shows that most of the rain accumulation over La Liboriana basin days before the time of the disaster was due to convective precipitation, with very intense spells. Figure 10a shows that a large part of the rainfall during the May 15 event was produced by convective precipitating clouds, and that during the most intense phase, lasting about 120 minutes, the area of the basin with convective precipitation was about the same as the stratiform-covered 340 region. By the end of the event, although the accumulation is low, there is a predominance of stratiform rainfall over the basin.

QPE and satellite precipitation
The second of the three events started during the night of May 16, and the average cumulative precipitation during period 2, in the basin, was around 45-50 mm (Figures 9c and d). The spatial distribution of rainfall during period 2 is the most homogeneous among the three events, typical of stratiform-dominated events covering the entire basin as observed in Figure   10d, and that was part of a broad cloud system that was covering a significant portion of western Colombia. By the end of 345 period 2, Figure 10d shows the presence of three short-lived convective cells in the basin covering an important fraction of the basin. Figure 11 shows weather radar horizontal reflectivity (Zh) retrievals depicting key moments of the evolution of this event over the entire 120 km radius area, and over La Liboriana basin. The black dots correspond to cloud-to-ground lightning. Rainfall during periods 1 and 2 increased the overall soil moisture in the basin, likely decreasing the magnitude of the infiltration rates (Penna et al., 2011;Zehe et al., 2010), hence increasing runoff, and the likelihood of flash floods occurrence (Wagner et al., 1999;Penna et al., 2011;Tramblay et al., 2012). Twenty hours after the second event, a third event characterized by training convection (deep convective cells organized such that they move repeatedly over the same area as described by Doswell et al. (1996)    The flooding was reported in the urban area of Salgar about 30 to 45 minutes after the occurrence of this intense event. The convective cores were immersed in a system of storms that reached the mesoscale (around 70 km of length), and was moving across the Department of Antioquia with a SE-NW direction. These events can be categorized as almost purely convective ( Figure 10e) with Zh reaching over 40 dBZ. It is relevant to note that the cumulative convective rainfall reported during the day of the disaster was not the highest among the three events, and its intensity was not exceptionally higher.
375 Figure 13 shows the vertical cross-sections of Zh, radial velocity (V r), differential reflectivity (Zdr), and polarimetric correlation coefficient (hv), across one of the convective cells that were occurring north of Salgar around the time of the disaster (02:00 LT, May 18, 2015) and was also immerse in system of the storms. An intense, purely convective storm can be observed reaching up to 14.5 km (Figure 13a), with the characteristic divergent pattern in the upper levels of the radial velocity field (Figure 13b). The intensity of the storm is associated with a core of high Zh values over a 7 km region (horizontal  extension). Figure 13c shows values of Zdr greater than 1, indicative of non-spherical particles likely composed of liquid droplets, associated with the high Zh values between 4 and 6 km height, and also aloft, over 10 km height towards the western side of the storm, where ice and snow particle formation is likely occurring. All of this occurs in a region with high particle homogeneity, regarding hydrometeor shape, characterized by high hv values (Figure 13d). The shape and values of the polarimetric variables observed in Figure 13 are highly indicative of the composition of the associated storms, showing 385 mostly liquid droplets, and with particle microphysics distributed homogeneously over the affected region. The assessment of the vertical structure of the convective cores is also relevant to evaluate their intensification potential: in cases like the one presented in Figure 13, the depth of the convective core, and the overall structure of the system, suggests a high potential for intense rainfall at the surface.   the steepest topography, above 3000 m.a.s.l., the intensities increase significantly, with cores exceeding 40dBZ. At around 01:30 LT, a precipitation system with high intensities approached Cerro Plateado, joining the first core, also generating a high-intensity nucleus that extended latitudinally from 5.93 • N to 6.06 • N (see Figure fig:PPI18M). These intense cores persist 400 close to the steepest and highest terrain for about 2 hours when the system starts to spread around, covering the basin with precipitating clouds of varying intensities (Figures fig:PPI18Mk and i). The intensification of the cores as they approach the topographic obstacle is evident in the three radar scanning tilts (Figures 14a to c). Subsequently, the system began to dissipate as it migrated out of the basin in a northwestern direction. According to Figure 14e, it is possible to identify that the approaching systems increased their intensity by at least 30dBZ when approaching the topographic barrier highlighting the importance of   Figure 9 for the IMERG (a, b, c, d, e, and f) and TRMM 3B42 (g, h, i, j, k, and l) products, correspondingly. Figure 15 shows the IMERG and TRMM 3B42 cumulative precipitation for the three events from May 14 to May 18, 2015, for the same regions and the same periods as in Figure 9. The areal extents shown, the 120 km radius area centered at the radar 410 site ("radar scale") and the La Liboriana basin, allow assessing the usefulness of both satellite products in risk management.

Evidence from satellite retrievals
At the radar scale, despite the apparent underestimation of peak values by the satellite retrievals, the overall structure of the rainfall distribution is captured skillfully by the satellite-based estimates, and in particular by the IMERG algorithm. In the latter case, IMERG captures the spatial structure of the cumulative precipitation on all three cases, including the location and areal extent of the events. On the other hand, at the "basin scale", none of the products capture the intense cores evident in the 415 QPE technique. This does not necessarily preclude the use of satellite information in risk analysis, but it does limit its direct use in flash flood warning systems. Alternatively, given that at the radar scale satellite information appears skillful, the spatial distribution of precipitation from, for example, IMERG, should be used as a two-dimensional probability density function (or mass spatial function) combined with downscaling schemes such as a multifractal framework (e.g. Deidda, 2000;Tao and Barros, 2010), to generate probabilistic higher-resolution precipitation fields conserving the original mass in the coarse scale. 420 Figure 16 shows the Hovmöller diagrams, at 6.00535ºN, of IMERG precipitation and GOES-13 Tb for the flash-flood triggering event during May 18, 2015. Figure 16a presents evidence that the IMERG dataset captured reasonably well the discerning among deep clouds and thin-high clouds using the infrared-thermal channels.

First-order hydrometeorological processes
One of the goals of this work is to assess the likelihood of occurrence of extreme events similar to the one, or to the ones, triggering La Liboriana flash flood. In other words, it is important to evaluate whether or not the characteristics of the May 18, 2015 flood were exceptional, and ideally, their recurrence rate. In a traditional sense, it would be desirable to estimate a return 430 period of the conditions that led to the la Liboriana flash flood. While the length of the historical radar QPE record is not enough for a robust estimation of the mentioned return period, the previous analysis together with first-order hydrometeorological considerations allow us to conduct a preliminary assessment of the exceptionality of the precipitation conditions associated with the event. Analyses in the previous sections suggest that (i) the spatial structure of precipitation relative to the basin's main geomorphologic features, (ii) the occurrence of multiple precipitation events in a relatively short timespan (3-4 days), and (iii) 435 the orographic enhancement of precipitation played a significant role in triggering the observed flash flood, suggesting that the traditional point rainfall return period estimation based on an intensity-duration-frequency curves (IDF), grossly misrepresents the observed hydrometeorological processes, even using areal transformations using reduction factors (e.g. Rodriguez-Iturbe and Mejía, 1974;Bacchi and Ranzi, 1996;Sivapalan and Blöschl, 1998;Veneziano and Langousis, 2005;Barbero et al., 2014).
In general, all approaches considering as a basis a spatially random distribution of precipitation over a specific basin would not 440 represent properly the observations. Considering the evidence suggesting that the occurrence of the 2015 La Liboriana flash flood is linked to multiple precipitation events in a few days, or in other words, that no single precipitation event was exceptionally large to generate the extreme event, Figure 18 shows the assessment of the combined role of precedent rainfall modulating overall soil moisture, and the intense precipitation during the event, by estimating the bivariate histogram of 48-hour (and 96-hour) cumulative precipitation record, but in particular, in the upper part of the basin, implying that for optimal risk management it is necessary to consider the spatial distribution of cumulative rainfall relative to the geomorphological features of the basin. In other words, while the  individual event on May 18 was not exceptional, the climatological anomalies were negative-to-normal, and the synoptic patterns around the extreme event were similar to the expected ones for the region, the combination of high rainfall accumulation as a result of successive precipitation events over the basin, followed by a moderate extreme event is unique in the available 475 observational record. The evidence also suggests that the 96-hour period, in this case, is more appropriate to analyze the extreme event (Figures 18c and d).

WRF forecasts
The use of accurate and skillful numerical weather prediction models is arguably one of the most promising strategies to 480 improve the lead times in flash flood forecasting schemes. In such a context, simulation and forecast skill refers to the ability of the model to capture large-scale moisture advection features responsible for the heavy rainfall to the region (e.g. Younis et al., 2008;Gochis et al., 2015). In regions with complex terrains, such as Salgar, in addition to the moisture advection, limited-area models are required to represent the processes leading to the observed orographic precipitation enhancement. In general, QPF precipitation forecasts of extreme events tend to underestimate the total rainfall amounts (Gochis et al., 2015), however, in 485 cases where the spatial distribution of precipitation is accurately anticipated, model output statistics (MOS) techniques could be used to bias-correct the QPF, for its use in flash flood likelihood assessment, providing valuable information for an optimal risk reduction.
A brief evaluation of the limited-area operational weather forecasts issued by SIATA, suggests that the model simulate reasonably well the main processes leading to the observed orographic enhancement that took place prior the La Liboriana 490 flash flood. In the case of the SIATA operational forecasts, which tends to underestimate total rainfall amounts, we have found that the 90th percentile of precipitation in a given area represents better the observed average (or median) precipitation over a speciric area, and we use it here as a simple quantile-to-quantile MOS technique. Figure 19 presents (Lin et al., 1983) schemes. Here we present the evaluation of the model forecasts using the Lin scheme given that, for the case of La Liboriana flash flood, the skill of the successive rainfall forecasts were considerably higher than using the other schemes. Figure 19a shows the comparison between the hourly time series of spatially-integrated rainfall forecasts and the precipitation obtained using the radar QPE described previously. The observed time series is obtained by averaging the QPE over the Salgar municipality, and the simulated time series selecting the 90th percentile of the precipitation forecasts over the same 500 region. In general, all the forecasts show heavy rainfall during the subsequent days matching the observations, especially the timing of the events. The May 14 forecast captures the timing of the series of all three events, but as lead-time is larger, it fails to capture the peak magnitude of the total amounts of precipitation. Similar performance is observed for all the different forecasts, with the May 16 forecast the most accurate capturing the rainfall event during May 17. Even though the event on May 18 was forecasted in all cases, the amount of precipitation was considerably less than the one observed. Figure 19b shows the total 505 amount of precipitation accumulated from the start of the forecast to May 19. In general, despite using the 90th percentile, the forecasts underestimate the observed rainfall amounts. In the May 14 forecast, the total precipitation corresponds to 160 mm over the Salgar region, and the radar observations show 390 mm (May 17 and 18 events were considerably underestimated).
Among all the forecasts shown in Figure 19a, the May 16 case shows a better agreement with observations as it captured the May 17 event, with a total accumulation of 120 mm compared to the observed 190 mm.
510 Figure 19c shows the spatial distribution of the May 14 1200 UTC 96-hour forecasted cumulative precipitation. While the forecast significantly underestimates the total amount of precipitation during the period in question (see Figure 9), the model The analysis of the spatio-temporal configuration of precipitation, identifying the main hydrometeorological factors control-555 ling the occurrence of extreme events and their likelihood of occurrence, is the foundation for the second risk management lesson. In the case of La Liboriana event, a series of intense storms associated with wind flow favored by a low-pressure system that was located over the Pacific coast of Colombia and Panama occurred days and hours preceding the disaster. The overall evidence of La Liboriana flash flood shows a definitive role of the multiple precipitation events in a relatively short period, and of their intensification as convective cores approached the steepest topography. There were three successive events generating 560 significant rainfall within La Liboriana basin: No single precipitation event was exceptionally large to generate the extreme event, but rather the combined role of precedent rainfall, and extreme hourly precipitation triggered the event. The first two events, that increased the overall soil moisture in the basin, were followed by a third event characterized by training convective elements towards La Liboriana with significant orographic enhancement with a preferential spatial distribution of the most extreme rainfall towards the upper basin, ultimately triggering the flash flood. Most of the rain over La Liboriana basin days 565 before the disaster was of the convective type, with very intense spells, generating high rainfall accumulations with the highest hourly cumulative precipitation linked to the steepest slopes in the basin due to the orographic intensification in the region.
Tactic risk management decisions would benefit from implementing, as a complement to flash-flood guidance tools, a realtime assessment of the bivariate distribution (and joint probability) of each ongoing event compared to the historical record, taking into account high one-hour moving cumulative rainfall conditioned to values of precedent N -hour moving cumulative