Elsevier Editorial System(tm) for Journal of Hydrology Manuscript Draft
Manuscript Number:
Title: Comparison of two systems for regional flash flood hazard assessment
Article Type: Research paper
Keywords: Flash floods; Early warning systems; Radar rainfall; Drainage network; Return period
Corresponding Author: Dr. Carles Corral,
Corresponding Author's Institution: Universitat Politècnica de Catalunya
First Author: Carles Corral
Order of Authors: Carles Corral; Marc Berenguer; Daniel Sempere-Torres; Laura Poletti; Francesco Silvestro; Nicola Rebora
Abstract: The anticipation of flash flood events is crucial to issue warnings to mitigate their impact. This work presents a comparison of two early warning systems for real-time flash flood hazard assessment and forecasting at regional scale. The two systems are based in a gridded drainage network and they use weather radar precipitation inputs to assess the hazard level in different points of the study area, considering the year return period as a good indicator of the flash flood hazard. The essential difference between the systems is that one is a rainfall-based system, using the upstream catchment-aggregated rainfall as the variable to determine the hazard level, while the other is a full system based on streamflows computed by a rainfall-runoff model. The comparison has been done for three rainfall events in the autumn of 2014 that resulted in severe flooding in the Liguria region (Northwest of Italy). The results obtained by the two systems have many similarities, particularly for larger catchments and for large return periods (extreme floods).
Suggested Reviewers: Katharina Liechti [email protected]
Lorenzo Alfieri [email protected]
Jonathan Gourley [email protected]
Kai Schröter [email protected]
Francesco Marra [email protected]
Highlights (3 to 5 bullet points (maximum 85 characters including spaces per bullet point)
Highlights Two flash flood early warning systems are compared analysing three extreme events. Hazard is assessed in terms of return period for the whole Liguria region. Weather radar rainfall is used as input for the warning systems.
Manuscript (double-spaced and continuously LINE and PAGE numbered) Click here to view linked References
1 Comparison of two systems for regional flash flood hazard assessment
2 Carles Corral, Marc Berenguer, Daniel Sempere-Torres
3 Center of Applied Research in Hydrometeorology. Universitat Politècnica de Catalunya,
4 Barcelona, Spain.
5
6 Laura Poletti, Francesco Silvestro, Nicola Rebora
7 CIMA Research Foundation, Savona, Italy.
8
9 Corresponding author: Carles Corral
11 CRAHI-UPC
12 Module C4 - S1, c/ Jordi Girona, 1-3. 08034 Barcelona
13
14 Abstract
15 The anticipation of flash flood events is crucial to issue warnings to mitigate their impact.
16 This work presents a comparison of two early warning systems for real-time flash flood
17 hazard assessment and forecasting at regional scale. The two systems are based in a
18 gridded drainage network and they use weather radar precipitation inputs to assess the
19 hazard level in different points of the study area, considering the year return period as a
20 good indicator of the flash flood hazard. The essential difference between the systems is
21 that one is a rainfall-based system, using the upstream catchment-aggregated rainfall as
22 the variable to determine the hazard level, while the other is a full system based on
23 streamflows computed by a rainfall-runoff model. The comparison has been done for three
24 rainfall events in the autumn of 2014 that resulted in severe flooding in the Liguria region
25 (Northwest of Italy). The results obtained by the two systems have many similarities,
26 particularly for larger catchments and for large return periods (extreme floods).
1 27
28 Keywords
29 Flash floods, early warning systems, radar rainfall, drainage network, return period
30
2 31 1. Introduction
32 Flash floods produce devastating effects worldwide. They affect steep and small to
33 medium catchments (up to few hundreds of square kilometres), and are typically induced
34 by extreme rainfall in the upstream area. The lapse time between strong rainfall initiates
35 (the cause) and the flood takes place (the effect) depends primarily on the features of the
36 catchment, ranging from the order of tens of minutes in the smallest catchments to a few
37 hours in the larger ones (Borga et al., 2008; Gaume et al., 2009; Marchi et al., 2010). In any
38 case, this reaction time is very short, frequently too short to activate the emergency
39 response and effectively prevent damages in human activities, properties and life. Thus,
40 anticipating this kind of hazard and get some extra minutes or hours for better
41 preparedness and response is crucial and a major challenge.
42 Among the problems encountered to extend this reaction time are the chaotic and
43 explosive nature of the rainfall events, usually characterized by convective systems
44 producing highly variable rainfall that are difficult to measure and forecast. Also, flash
45 floods affect places at regional scale that are often poorly or very poorly gauged at
46 catchment scale. That is, the intrinsic difficulties to apply a rainfall runoff model would be
47 amplified. For this reason, the high space and time resolution of radar rainfall
48 observations provide relevant information to produce both quantitative precipitation
49 estimates (QPE) and forecasts (QPF), which are fundamental inputs of a quality flash flood
50 early warning system.
51 Flood early warning systems provide real-time information about flood forecasts, and
52 consequently a planned warning to support the response can be issued. In order to be
53 useful for the practitioner, the magnitude of the flood is commonly related to the
54 exceedance of a particular threshold, for instance the river bank overflow or a statistical
55 frequency of occurrence. Ideally, flood forecasts should be directly associated to the flow
56 discharge in a specific section, but the limitations to implement an operational rainfall-
3 57 runoff model over large regions, or the difficulties to obtain reliable flow thresholds at the
58 same scale, have yield to other approaches. Since the upstream catchment rainfall is the
59 most important forcing factor to produce flow in a point, this value (spatially averaged)
60 can be used as the determining variable, leading to a system operationally more simple, at
61 the expenses of (in theory) loss of reliability. In a paper reviewing recent advances in flash
62 flood forecasting, Hapuarachchi et al. (2011) distinguished the two approaches described
63 above as the commonly used when QPE and QPF are available, naming them as the flow
64 comparison method and the rainfall comparison method, respectively.
65 The simplicity of the rainfall based methods does not only concern their implementation,
66 data needs and computer requirements. Any internal analysis or post-processing of the
67 results is in general simpler too. For instance, uncertainty analysis is limited to that of
68 rainfall, and forecasted rainfall is probably the main source of uncertainty in hydrological
69 modelling. Another aspect that highlights the interest of the rainfall based methods was
70 described in the development of the Gradex method by Guillot and Duband (1967), who
71 found that the gradient of the statistical distribution of the yearly maximum discharge
72 tends to follow asymptotically that of the yearly maximum daily rainfall for high return
73 periods. Then, when dealing with extreme floods, rainfall based methods seem to be
74 appropriate.
75 The Mediterranean coast is an area particularly prone to damages produced by flash
76 floods. Large rainfall amounts produced by the combination of the warm sea and the
77 abrupt orography is the main triggering cause, but also steep and narrow streams in
78 catchments where population and urban development have increased in the recent past
79 decades, often without an adequate flood urban planning. The Liguria region in Northwest
80 Italy shares the previous features of a Mediterranean flash flood prone area. In fact, its
81 location in the Gulf of Genoa makes the region particularly affected by intense rainfall.
4 82 A flood warning system is operating at the Regional Environmental Protection Agency of
83 Liguria (ARPAL), monitoring the entire region (Silvestro et al., 2011; Silvestro and Rebora,
84 2012). This system is based on the application of a distributed grid-based rainfall-runoff
85 model, and thus hazard assessment is based on a flow comparison method. Considering
86 the extensive spatial coverage of the outputs (at grid cell scale), it can be seen as a "full"
87 flood warning system applied at regional scale. The purpose of this study is to compare the
88 outputs of this flow based warning system against a rainfall based system which is similar
89 in structure and provides the same spatial coverage. This rainfall based flood warning
90 system is operational in the Spanish Regional Water Agency of Catalonia (ACA), and with
91 little effort it is transportable to other regions (Corral et al., 2009; Alfieri et al., 2011, 2017;
92 Versini et al., 2014). The comparison between the two systems provides the opportunity
93 to analyse their similarities and the differences. Among other conclusions, this comparison
94 should allow us to characterize the effect of the simplifications of a rainfall based warning
95 system compared to a flow based warning system.
96 Both systems provide hazard assessment in the form of frequency of occurrence (or
97 return period), and the comparison is made in these terms. We have analysed three
98 important flood events that affected the Liguria region in the autumn of 2014. For
99 consistency purposes, rainfall inputs are the same for both systems (radar rainfall
100 estimates during the selected events), and rainfall and flow thresholds are based on the
101 same climatological dataset. Although both systems are designed to incorporate rainfall
102 forecasts (QPF), for the sake of clarity only QPE are used in this study, thus concentrating
103 the analysis on the intercomparison.
104 The paper is structured as follows: Section 2 and section 3 introduce the rainfall based and
105 the flow based flood warning systems, respectively. Section 4 describes the study area,
106 and Section 5 deals with the events and rainfall data. Section 6 is about the results. Finally,
107 some concluding remarks are pointed out in Section 7.
5 108
109 2. Description of the rainfall based flood warning system (RFWS)
110 The rainfall based flood warning system (RFWS) is based on estimating the probability of
111 occurrence of the observed upstream averaged catchment rainfall (referred to as
112 catchment-aggregated rainfall) by comparison with the thresholds obtained from a
113 climatology of the same variable. The catchment-aggregated rainfall is computed at cell
114 scale over a gridded drainage network obtained with the D8 direction approach,
115 considering different rainfall accumulation windows. The particularity of this system is
116 that fixes the duration that characterizes the catchment for warning computations as the
117 concentration time. Thus, each cell representing an upstream catchment is related to an
118 individual concentration time [currently the Kirpich (1940) formula is used], and
119 computations are made for this duration.
120 The rainfall thresholds used for comparison have been obtained by processing the
121 Intensity-Duration-Frequency (IDF) curves based on rainfall climatology. Using rainfall
122 series from 125 raingauges, Boni et. al. (2007) applied a regional approach to derive
123 rainfall frequencies over Liguria, choosing the two components extreme value (TCEV)
124 distribution function for the parent distribution of the rainfall growth factor. These curves
125 are spatially distributed over the grid, and they are referred to as point IDF, IP(Di,Fj).
126 Operationally, each rainfall duration and frequency are related to a rainfall intensity map.
127 Subsequently, catchment-aggregated rainfall is computed for each duration and for each
128 frequency, obtaining catchment IDF maps, referred to as IC(Di,Fj). But in this case, knowing
129 that the IP(Di,Fj) integrated over a surface should be reduced by an areal factor (see for
130 example Pavlovic et. al., 2016), catchment aggregation means not only the spatial
131 averaging but also the application of a simple areal reduction factor given by Ka=1-
132 log10(S/15) (which is used in Spain for the design of discharge structures, MOPU, 1990)
133 where S is the surface catchment in km2.
6 134 If an important reservoir exists upstream of a cell, then the catchment is split at the dam
135 level and the upstream catchment is not included for catchment-aggregated computations
136 (thus assuming total regulation).
137 Then, given the QPE (and the QPF if available) at a specific operation time, the catchment-
138 aggregated rainfall is computed at cell scale over the drainage network, considering an
139 accumulation window equal to the catchment concentration time (operationally, the
140 system only works with previously specified durations, and it uses the duration value
141 closest to the concentration time). Comparing this against catchment-aggregated rainfall
142 thresholds for this specified duration, the frequency (representing the probability to
143 exceed a given value) is derived over the forecasting window, normally expressed as a
144 return period (average time between two occurrences).
145 This procedure is simple and easy to implement at regional scale, needing only
146 information about rainfall climatology to build the IDF curves, and a terrain elevation DEM
147 to derive the drainage network and the concentration time map.
148
149 3. Description of the flow based flood warning system (FFWS)
150 The system existing in the Liguria region is based on the application of the Continuum
151 rainfall runoff model (Silvestro et al., 2013, 2015), a continuous distributed hydrological
152 model that relies on a morphological approach, exploiting the information derived from
153 the drainage network (Giannoni et al., 2005). The rainfall runoff model has been conceived
154 to be a compromise between models with a strong empirical connotation, which are
155 simple to implement but far from reality, and complex physically-based models which try
156 to reproduce the hydrological processes with high detail but introduce complex
157 parameterizations and consequent uncertainty and lack of robust parameters
158 identification.
7 159 This system uses a D8 gridded drainage network (the same as the RFWS). Infiltration and
160 subsurface flows are described using a semi-empirical, but quite detailed, methodology
161 based on a modification of the Horton algorithm (Bauer, 1974; Diskin and Nazimov, 1994;
162 Gabellani et al., 2008); it accounts for soil moisture evolution even in conditions of
163 intermittent and low-intensity rainfall (lower than the infiltration capacity of the soil). The
164 energy balance is based on the so called ‘‘force restore equation” (Dickinson, 1988) which
165 balances forcing and restoring terms, with explicit soil surface temperature prognostic
166 computation. Vegetation interception is schematized with a storage which has a retention
167 capacity (Sv) estimated with the Leaf Area Index data, while the water table and the deep
168 flow are modelled with a distributed linear reservoir schematization and a simplified
169 Darcy equation.
170 The surface flow schematization distinguishes between hillslope and channelled flow by
171 means of a morphologic filter defined by the expression ASk=C, where A is the contributing
172 area upstream of each cell, S is the local slope, and k and C are constants that describe the
173 geomorphology of the environment (Giannoni et al., 2000). The overland flow (hillslopes)
174 is described by a linear reservoir scheme, while for the channelled flow the kinematic
175 wave approach is applied (Wooding, 1965; Todini and Ciarapica, 2001).
176 Continuum has six parameters that require calibration at the basin scale: two for the
177 surface flow (uh and uc), two for the subsurface flow (ct and cf) and two for the deep flow
178 and watertable (VWmax and Rf) processes. At the regional scale of Liguria, the model was
179 calibrated using observed flow data from 11 level gauge stations (Davolio et. al., 2017), by
180 means of a multi-objective function based on the Nash-Sutcliffe Efficiency and the relative
181 error on high flows. The six model parameters are assumed to be uniform at catchment
182 scale, and average parameters obtained in calibrated basins are imposed where no
183 observations are available.
8 184 The flow thresholds associated to the different probabilities of occurrence (return period)
185 over the region of Liguria were obtained from a statistical regional analysis (Boni et. al.,
186 2007). Flood peaks were obtained over the region applying a simple rainfall-runoff model
187 to each cell of the drainage network, using rainfall hyetographs given by a particular
188 Alternating Block Method which maximise the peak response. These hyetographs were
189 built based on the same rainfall frequencies (IDF information) that are used as thresholds
190 in the RFWS. Boni et. al. (2007) verified that the peak flow annual maxima provided by
191 this method are consistent with the regional flow analysis made using the historical data
192 from 33 stream flow stations.
193 Once these thresholds are derived over the working drainage network, a simple
194 comparison against the flow simulations from the Continuum model (during the
195 forecasting window) allows to estimate the return period at each cell of the region.
196
197 4. Study area
198 The Italian region of Liguria is a narrow strip of land about 250 km long and 20–30 km
199 wide with a surface area of about 5400 km2. It has very few flat areas and is mainly
200 covered by forest. Most of the catchments have their outlet directly in the Mediterranean
201 Sea and, because of the mountainous characteristics of the region, the main urban areas
202 and towns have been established along the coast, often at the mouth of a river. Many
203 basins are small, with an area of less than 100 km2, and prone to flash floods. Only a few
204 catchments have a drainage area over 200 km2, but in these cases their response times are
205 also very short, of just a few hours.
206 The Liguria region has a real-time hidrometeorological network that provides a detailed
207 set of meteorological variables. There are about 120 automatic weather stations, each one
208 equipped with a rain gauge. The region is also covered by a Doppler polarimetric C-band
9 209 radar, located on Mount Settepani at an altitude of 1386 amsl. The rainfall data used in the
210 present study corresponds to QPEs obtained from the observations of the Settepani radar.
211 These rainfall estimates were produced at 1 x 1 km2 and 10 minutes resolution, applying
212 an algorithm that selects the optimal relationship between rainrate and dual polarisation
213 radar variables based on a flowchart decision tree (Silvestro et al., 2009). The same QPEs
214 have been used as inputs to both the RFWS and FFWS.
215 From a Digital Elevation Model (DEM) of the region with a resolution 0.005 degrees both
216 in longitude and latitude, flow directions were identified based on maximum local slope
217 following the D8 approach (each cell is given a unique direction from its 8 neighbours,
218 O'Callaghan and Mark, 1984), and the drainage network is defined at the DEM resolution
219 (note that at this latitude 0.005 degrees means around 400 m in the WE direction, and
220 around 550 m in the SN direction).
221 Figure 1 shows the original DEM grid extension where the RFWS and the FFWS have been
222 applied, covering completely the Liguria region. A sub-area around Genoa covering the
223 zones most affected by the analysed events has been selected for displaying the results.
224 The most important streams located in this sub-area are: the Bisagno creek (crossing the
225 city of Genoa), the Polcevera creek (with its outlet at the Genoa port), the Entella river
226 (formed by the junction between the Lavagna and the Sturla rivers at Carasco), and the
227 Orba and the Scrivia rivers (these two flowing to the North to the Po plain). Six control
228 points have been selected in these streams, allowing to analyse the temporal evolution of
229 the results at local scale. Three of these selected points correspond to the location of
230 existing stream gauges: Genoa in the Bisagno creek (Passerella Firpo), Tiglieto in the Orba
231 river and Panesi in the Entella river (see bottom panel of Fig. 1).
232
233
10 234 5. Analysed events
235 The autumn of 2014 was particularly rainy in Liguria. A concatenation of rainfall events
236 affected the region causing important flash floods in several locations. We have selected
237 three events that produced several local floods, being the events of 09 October, and 10 and
238 15 November the scope of this paper.
239 The event of October was particularly devastating and was in-depth described in a paper
240 concerning the flood of the Bisagno creek (Silvestro et al., 2016). Daily accumulations over
241 200 mm were observed in a large area between Genoa and La Spezia (about 80 km south-
242 east of Genoa, see Figure 1). Rainfall intensities were very extreme (for example, a
243 raingauge located near Genoa recorded more than 130 mm in one hour). The observed
244 peak flow in the Bisagno creek (90.9 km2) was around 1100 m3/s, and about 800 m3/s in
245 the Entella river (370.6 km2). Flash floods caused major damages over the region, one
246 person lost his life and material damages were estimated in around 100 Million Euros.
247 Compared to the event of 09 October, the other two selected events are of lesser
248 importance, and the spatial affection of flooding was also smaller. The event of 10
249 November mainly affected the Entella river (peak flow about 1300 m3/s), and small
250 overflows were reported in the village of Carasco. On the event of 15 November the most
251 affected zone corresponds to the coast in the West of Genoa, being reported local
252 overflowing in the Polcevera creek around Pontedecimo, and also local disasters in other
253 creeks nearby. The Orba river also experienced a significant increase in discharge (peak
254 flow about 500 m3/s at the stream gauge of Tiglieto), but without overflooding. A
255 summary of the effects produced in these three events can be read in
256 https://it.wikipedia.org/wiki/Lista_di_alluvioni_e_inondazioni_in_Italia.
257 As a way to summarize the quality of the radar based QPEs used in the study, Figure 2
258 shows a comparison of the event total rainfall between raingauge observations and the
259 collocated radar estimates (one scatter plot for each event). All values correspond to
11 260 rainfall accumulations in 48 hours. For the comparison, only raingauges located within 80
261 km from the radar location have been selected.
262 The regression line in the event of 09 October shows a very small bias. In contrast, the two
263 events of November present more scatter and the regression lines differing from the
264 perfect fit (regression slope about 0.65), showing a clear trend to underestimation. The
265 apparent differences between the first event and the other two can partly be explained by
266 the spatial affection of the rainfall event, that in the first case was very restricted and most
267 of raingauges registered less than 10 mm.
268 These results show that these radar QPEs are already affected by several error sources,
269 particularly range dependent (due to the Vertical Profile Reflectivity, beam broadening or
270 path attenuation, for example) and also related to the complex orography of the region
271 (ground clutter, beam blocking). Although the quality of these products might affect the
272 performance of the two flood warning systems (in ARPAL operation they are merged with
273 raingauge data), in this study, focused on the intercomparison, radar-only QPEs have
274 enough quality to provide a valid basis to evaluate differences and similarities of the
275 simulations.
276
277 6. Results
278 The results obtained applying the two flood warning systems are presented here. First
279 sections (6.1-6.3) focus separately on each event, summarizing the results in a map of the
280 selected sub-area and showing the results obtained in some control points. Section 6.4
281 presents the comparison between the two systems globally for the entire region, taking
282 into account all the points of the drainage network and analysing the impact of several
283 factors.
284
12 285 6.1. Event of 09 October 2014
286 Figure 3 summarizes the results obtained in the Event of 09 October 2014, which is by far
287 the most important of the three cases. The figure shows the maximum return period
288 obtained over the 48-hour time window by each flood warning system (excluding the
289 pixels with a drainage area of less than 4 km2). Rainfall accumulation maps over the entire
290 region are also included.
291 Although there was significant rainfall only in part of the region, the 48-hour rainfall
292 accumulation exceeds 200 mm over a very large area, and even 600 mm (from radar) are
293 exceeded over an area at the north-east of Genoa. The maximum accumulation observed
294 by a raingauge in this period was in Torriglia (about 8 km East of Montoggio) with a total
295 of 476 mm.
296 Using the two systems, the maximum return periods were achieved in the Bisagno creek,
297 the Entella river and the Scrivia river (this last one flowing to the north). Comparing the
298 results provided by both systems, the return periods obtained by RFWS are in general
299 smaller. Very important return periods were obtained by the FFWS in the medium course
300 of the Bisagno creek (over 200 years), and this is where the differences with the RFWS are
301 more apparent (return periods are below 100 years).
302 The results of the event are also presented in terms of the time series of estimated return
303 period obtained at 10-min temporal resolution in the most interesting control points
304 (shown with thicker circles in Fig. 3). In addition, when a stream gauge exists at the place,
305 the warning level associated to the observed flow discharge (1-hour resolution) is also
306 included.
307 Figures 4-6 show the estimated return periods obtained in three control points. Results in
308 the Genoa control point (Fig. 4) show very similar results between the two systems, in
309 particular the peak (both around 125 years), and they provide a very accurate agreement
13 310 with the return period estimated from flow observations. The coarse time resolution given
311 by the observed flow (1 hour) does not allow any conclusion about the performance of the
312 systems during the rising limb (with 10-min resolution). In the case of the Entella river at
313 Panesi (Fig. 5), the warning level provided by FFWS (63 years) is quite higher than that of
314 the observed flow (5 years), and the RFWS (7 years) provides a more reliable estimate.
315 Finally, high warning levels are also obtained in the control point of Montoggio in the
316 Scrivia river (Fig. 6), being the return period of the FFWS (124 years) higher than that of
317 the RFWS (92 years). Results obtained in Genoa and Montoggio (return period around 100
318 years) seem in accordance with the overbank inundations actually reported in these
319 places.
320
321 6.2. Event of 10 November 2014
322 The rainfall accumulation map for this event shows the larger amounts distributed around
323 the coast, with two maxima over 200 mm (Figure 7). But only in the Entella river and its
324 tributary Sturla both systems issued significant warnings. Comparing the results provided
325 by the systems by eye, FFWS provides in general higher return periods, with maxima over
326 30 years.
327 Figures 8 and 9 show the results obtained in the control points around the Entella river.
328 Some differences arise between the results obtained in the two control points. In Panesi
329 (on the Entella river with an area of 371 km2) the FFWS provides warnings over 10 years
330 (16), which is quite similar to that provided by the observed flow (about 13 years), while
331 the maximum provided by RFWS is 8 years. But in the case of the Carasco control point
332 (outlet of the Sturla river just before flowing into the Entella, with 132 km2), the results
333 are more similar (particularly the rise and the time to peak), although the RFWS peak (35
334 years) is higher than that of FFWS (25 years). Actual overflows were reported in the
335 village of Carasco (confluence of the Lavagna and Sturla rivers), while downstream in the
14 336 Entella river there is no proof of overbank inundations. The return periods obtained by
337 both warning systems are in accordance with these reported inundations.
338
339 6.3. Event of 15 November 2014
340 As the other events, the event of 15 November 2014 also shows a rainfall pattern around
341 the coast of Liguria (Figure 10), but in this case concentrated mainly in the western part of
342 the region. The most significant signal has been obtained in the Polcevera creek, with
343 maximum return periods over 100 years estimated using the two warning systems,
344 although the river courses where these maxima are located differ from one system to the
345 other.
346 In this event, only the warning levels obtained in the control points of Pontedecimo
347 (Polcevera creek) and Tiglieto (Orba river) are interesting (Figures 11 and 12). In the first
348 case, there are two rainfall peaks having a time lapse of about two hours. While the FFWS
349 reflects these peaks with warning levels over 10 years (22 years in the second peak), the
350 RFWS only provides significant warning levels in the second phase (16 years), showing
351 certain delay with respect to the FFWS outputs. It is interesting to note that during this
352 event, the Polcevera creek in Pontedecimo and downstream actually suffered minor
353 overbank flooding, which seems in accordance with the estimated maximum return
354 periods.
355 In the case of the Tiglieto control point (Fig. 12), the outputs from both systems are quite
356 similar around the peak (19 years for FFWS; 16 years for RFWS). And both systems are
357 quite close to the warning level derived from the observed flow available at this place (17
358 years).
359
15 360 6.4. General comparison analysis
361 The analysis made at the control points allows to clearly compare the outputs of the two
362 systems, but it is limited to these specific places. The results obtained at the control points
363 have shown, in general, some agreement between the two systems, particularly the
364 evolution during the event. To provide a more general vision, the maximum return periods
365 obtained by both systems have been selected at pixel scale over the entire domain, and
366 they have been displayed in a scatter plot. From the total number of pixels of the domain,
367 only those pertaining to the main drainage network (having a drainage area greater than 4
368 km2) have been selected.
369 The idea of this analysis is to study whether the direct contrast provided by these scatter
370 plots (Figures 13-15) provides information about the differences and similarities of the
371 two warning systems, and if they can be related to other factors. In particular, the analysis
372 focuses in the catchment extension (drainage surface); the catchment itself (since the
373 rainfall-runoff model integrated into the FFWS is calibrated at catchment scale); and the
374 initial moisture state of the catchment (since this is a determining variable to generate
375 surface runoff from rainfall).
376 Thus, from the total number of points present in this analysis (6277 points in the main
377 drainage network), we have grouped them in three ways: (1) their drainage surface,
378 having more or less than 50 km2 (two classes); (2) if they belong or not to one of the five
379 main catchments in the selected sub-area of Liguria (Bisagno, Scrivia, Entella, Polcevera
380 and Orba); and (3) the average soil moisture of the upstream catchment at the beginning
381 of the event, simulated at cell scale by the Continuum model and expressed as the
382 proportion of the model soil reservoir that is full (referred to as relative humidity, RH,
383 with 4 classes).
384 This last point is more complex than the other two, and needs the way to determine the
385 time related to the beginning of the rainfall event to be defined. But frequently the rainfall
16 386 events consist of a concatenation of events with high spatial variability, and the definition
387 of the beginning of the event for a large area can be difficult and affected by some
388 subjectivity. Since this analysis is focused on the maxima, we have considered that the
389 period that effectively affects the peak results corresponds to the previous period with a
390 time duration given by the concentration time of the catchment (in consistency with the
391 background of the RFWS). Thus, since peak values are achieved at different times and
392 concentration times varies at each point of the drainage network, each point has been
393 associated to a different state of the soil reservoir grid simulation provided by the
394 Continuum model. For conservative purposes, peak values correspond to both RFWS and
395 FFWS results, keeping the earlier time among the two peaks.
396 Although the results of the three events are included in this section, the events remain to
397 be analysed separately (Figures 13-15). The scatter plot axes are scaled in a double
398 logarithm way by means of the variable u, known as the Gumbel reduced variable (Eq. 1),
399 where T is the return period (in years).
400 (Eq. 1)
401 The upper-right scatter plots of Figures 13-15 do not seem to show any clear relationship
402 with the location (see catchment grouping). In fact, the results obtained in any of the five
403 analysed catchments present a high scatter. In contrast, the responsible of this scatter
404 seems to be mainly the points having an associated catchment surface lesser than 50 km2
405 (upper-left scatter plots). If we keep only the points with a catchment surface greater than
406 this value, a clear trend is observed: in general, RFWS warning levels are smaller than
407 those obtained by FFWS, particularly for small return periods (almost systematically),
408 while for greater return periods the two warning systems present a better agreement
409 (although the general trend remains to be that RFWS return periods are a bit smaller).
410 This result is in accordance to the hypothesis highlighted in the introduction that, for small
17 411 probabilities of occurrence, discharge and rainfall probabilities tend to follow the same
412 gradient (for maximum yearly values).
413 The analysis grouping the pixels according to its upstream catchment initial soil moisture
414 (Figures 13-15 bottom) shows a logical evolution after these important rainfall events
415 over a period of less than 25 days: while many catchments are relatively dry at the
416 beginning of the first event (October 2014), they are almost completely wet at the
417 beginning of the third event. In any case, the event of October was characterised by a
418 strong rainfall persistence over the affected catchments, and the state of the soil reservoir
419 of these catchments was very wet prior to the main rainfall event (RH higher than 0.70).
420 It is interesting to note that, as a general fact, high return periods are not obtained in
421 catchments that are dry at the beginning of the event. There are only a few catchments
422 having a relative humidity of less than 0.70 where a return period of 8 years is achieved
423 (u>2). In fact, only a few points having an initial RH smaller than 0.50 even appear in the
424 scatter plots (mainly in the event of 09 October, Figure 13 bottom panel), meaning that at
425 least one of the systems did not obtain a significant return period, and thus it is not
426 possible to derive any conclusion for dry conditions. As it has been stated, the general
427 trend is that FFWS peaks are higher than RFWS peaks, and this is apparent for the points
428 having the largest values of RH (0.70-1.00), and also for medium values of RH (0.50-0.70).
429 In order to somewhat quantify the similarities of the two systems, the Critical Success
430 Index (CSI, Schaefer, 1990) between the FFWS and RFWS is computed for the peak values
431 over certain thresholds (return periods higher than 2, 10 and 20 years, respectively). The
432 CSI has been evaluated over the whole area, and also following the classification with the
433 drainage area, catchment pertaining and initial soil moisture. The results are summarised
434 in Table 1. One can appreciate high CSI values in the catchments that were affected by
435 extreme floods, as it is the case for Bisagno, Entella and Scrivia in the event of 09 October
436 and for Polcevera in the event of 15 November. When CSI is computed for extreme floods
18 437 (over 10 years return period), then it is clear that better agreement is obtained in the
438 catchments with a drainage area over 50 km2 than in smaller catchments. And when
439 looking at the influence of the initial soil moisture, better agreement is obtained in the
440 catchments with RH values higher than 0.70 than in drier catchments (in general for all
441 the different return period thresholds tested), but it is probable that the CSI obtained for
442 RH values lesser than 0.70 would be contaminated by the fact that there are too few
443 effective points with a related peak over the threshold (for the FFWS or the RFWS),
444 making the CSI computation very sensitive.
445
446 7. Discussion and conclusions
447 The comparison of the two flood warning systems in three events suggests to ask whether
448 the obtained results are expected or not. The comparison made at specific control points
449 reveals a quite good agreement between the two systems, and contrasting against flow
450 observations, it is not always clear that the flow based flood warning system (FFWS)
451 performs better than the rainfall based system (RFWS). In any case, observed flow data
452 have been only available at 3 stations, and the 6 selected control points are not
453 representative enough of the territorial extension of the study area (they have in common
454 a relatively large drainage surface). In addition, rainfall estimates from radar are affected
455 by some error sources, and they suffer from certain bias (particularly in the events of
456 November, Fig. 2). Thus, we assume that the results obtained at the control points are only
457 descriptive and qualitative.
458 When comparing both warning systems over the entire domain (the Liguria region), a
459 quite spread behaviour is obtained. Contrasting the maximum return period obtained at
460 each pixel by each one of the systems in a scatter plot, the agreement between FFWS and
461 RFWS can be considered low. As a general trend, return periods provided by RFWS are
462 lower than those provided by FFWS, although the agreement is clearly closer when
19 463 dealing with extreme floods, particularly for larger catchments (above 50 km2). And
464 surprisingly the effect of the soil moisture of the catchment does not show a clear
465 influence on the results. Then, if the initial soil moisture is not a determining factor to
466 explain the differences between the two systems, it should be the own dynamics of the
467 Continuum model during the rainfall event.
468 The RFWS explored in this paper has similarities with the European Precipitation Index
469 EPIC presented by Alfieri et al. (2011). This index is a sort of catchment-aggregated
470 rainfall normalised by the mean of annual maxima, computed for different rainfall
471 durations (6, 12 and 24 h) and keeping the value that provides the maximum index. In a
472 recent paper, Alfieri and Thielen (2015) made a similar comparison as that shown in
473 Figures 13-15, contrasting the EPIC index (KPmax) against a parallel KQ index obtained from
474 the simulated discharge (using a rainfall runoff model) normalised by the mean of annual
475 maximum discharges. Their results showed that the rainfall index (EPIC) generally
476 provided higher return periods than the flow index (KQ), over the full range of results
477 (return periods ranging from 1 to 200 years). Thus, these results were notably different
478 than that of the present paper, where the general trend is that RFWS provides lower
479 return periods in front of FFWS. This can be probably related to the fact that the
480 catchments affected by heavy rainfall were quite wet prior to the beginning of the main
481 rainfall event, in the present study.
482 The performance of the RFWS used in this study is quite dependent to the duration used
483 to accumulate rainfall. Since this duration is related to the concentration time of the
484 catchment, the formulation used to obtain this parameter seems quite critical. In fact, the
485 concentration time responds to a conceptualisation of the catchment, and the derived
486 existent formulations are empirical. As stated by Grimaldi et al. (2012), a universally
487 accepted working definition of this parameter is currently lacking, and available
20 488 approaches for the estimation of the time of concentration may yield numerical
489 predictions that differ from each other by up to 500%.
490 As an exercise, in addition to the formulation used to compute the concentration time (the
491 Kirpich method, 1940), we have tested the Témez method (commonly used in Spain for
492 discharge structures design, MOPU, 1990). Both need the same information, the length of
493 the upstream longest course L (in km) and the average slope S of this course, following a
494 generic formula given by the Eq. 2, being c=0.0664; =0.77; =0.385 for Kirpich method,
495 and c=0.3; =0.76; =0.19 for Témez method. Given the values usually found in
496 catchments like those treated in this study, the second method provides concentration
497 times that are of the order of the double than the former.
498 (Eq. 2)
499 Considering FFWS as a reference, it has been seen that the selection of the Kirpich
500 concentration time produces better agreement in terms of the warning level peak, and
501 particularly in terms of timing. As an example, the Figure 16 shows the results in the
502 control point of Genoa for the event 09 October 2014. In this case, the return period peak
503 obtained with the alternative method is almost the same than that obtained using the
504 Kirpich method, although in other cases (not shown here) there is a substantial variation
505 in the peak. During the recession there is an extension of high warning levels during about
506 3 hours. For operative purposes, providing accurate warnings during the recession is of
507 lesser importance. In contrast, accurate warnings during the rising limb is crucial in flash
508 flood forecasting, and in the case of the Figure 16 a systematic delay exists of the order of
509 10 to 20 minutes in relation to the results obtained using the Kirpich concentration time.
510 In relation to the FFWS results (the reference), another delay of 10 to 20 minutes should
511 be added. Then, an adequate selection of the concentration time used in the RFWS seems
512 to be an important decision. Our feeling is that this characteristic time does not have to be
513 strictly referred to the formal definition of the concentration time (the time that lasts a
21 514 water drop located in the more distant point to reach the outlet), but to a time that allows
515 most of the catchment to contribute. In this sense, the Kirpich method, which provides
516 quite reactive times as a concentration time, has resulted to be a good estimate in the
517 study domain.
518 The rainfall thresholds used in the rainfall based warning system (RFWS) have been
519 obtained from raingauge historical series, having the IDF curves at these specific places.
520 Although they can provide good estimates at local scales when quality and long series are
521 available, several problems arise when dealing with catchment-aggregated rainfall that
522 surely affect the related uncertainty: the need to spatially interpolate rainfall and the need
523 to apply an "ad-hoc" surface reduction factor. In this sense, when a long series of radar
524 QPEs are available, they are an opportunity to generate catchment IDF curves (see
525 Panziera 2016), and this will end in a more consistent approach for the RFWS.
526 In this study, we had not enough quality information to quantify the real magnitude of the
527 analysed flood events in terms of return period at the regional scale (just a few streamflow
528 gauges). Then, an accurate analysis of the reliability of the warning systems is not possible,
529 and it has not been the object of this intercomparison to elucidate whether one flood
530 warning system is better to the other one. However, this study has been useful to check
531 whether a rainfall based flood warning system can be reliable and useful. It can be stated
532 that if a distributed rainfall runoff model is calibrated for a region (with a satisfactory
533 index of agreement against flow observations), if flow thresholds are well defined over the
534 whole region and if computer time requirements are not a major problem, there should be
535 no doubt: the recommendation is to use a flow based warning system. If it is not the case
536 (and unfortunately this is quite frequent), we think that a rainfall based flash flood system
537 is an option that can be almost as reliable as a flow based system, in particular for larger
538 catchments (larger than 50 km) and when dealing with extreme events (return periods
539 over 20 years). This is reinforced by the fact that the major source of uncertainty in a flood
22 540 warning system is due to rainfall estimates, particularly forecasting rainfall, and this
541 uncertainty will affect similarly any of these warning systems.
542
543 Acknowledgements
544 This work has been done in the framework of the European Projects RISC-KIT (FP7-ENV-
545 2013-603458) and ANYWHERE (H2020-DRS-1-2015-700099), and of the Spanish Project
546 FFHazF (CGL2014-60700R).
547
548
549 References
550 Alfieri, L., Velasco, D., Thielen, J., 2011. Flash flood detection through a multi-stage
551 probabilistic warning system for heavy precipitation events. Adv. Geosci. 29, 69 – 75,
552 doi:10.5194/adgeo-29-69- 2011.
553 Alfieri, L., Berenguer, M., Knechtl, V., Liechti, K., Sempere-Torres, D., Zappa, M., 2017. Flash
554 Flood Forecasting Based on Rainfall Thresholds. In: Handbook of Hydrometeorological
555 Ensemble Forecasting. Duan, Q. , Pappenberger, F., Thielen, J., Wood, A., Cloke, H. L.,
556 Schaake, J. C. (Eds.), Springer Berlin Heidelberg, doi:10.1007/978-3-642-40457-3_51-1.
557 Alfieri, L., Thielen, J. 2015. A European precipitation index for extreme rain-storm and
558 flash flood early warning. Met. Apps, 22, 3–13. doi:10.1002/met.1328.
559 Bauer, S., 1974. A modified Horton equation during intermittent rainfall. Hydrol. Sci. Bull.
560 19, 219–229.
23 561 Boni, G., Ferraris, L., Giannoni, F., Roth, G., Rudari, R., 2007. Flood probability analysis for
562 un-gauged watersheds by means of a simple distributed hydrologic model. Adv. Water
563 Resour. 30 (10), 2135–2144. doi:10.1016/ j.advwatres.2006.08.009.
564 Borga, M., Gaume, E., Creutin, J.D., Marchi, L., 2008. Surveying flash flood response: gauging
565 the ungauged extremes. Hydrol. Process. 22 (18), 3883–3885. doi:10.1002/hyp.7111.
566 Corral, C., Velasco, D., Forcadell, D., Sempere-Torres, D., Velasco, E., 2009. Advances in
567 radar-based flood warning systems. The EHIMI system and the experience in the Besos
568 flash-flood pilot basin. In: Flood Risk Management: Research and Practice. Samuels, P.,
569 Huntington, S., Allsop, W., Harrop, J. (Eds.), Taylor & Francis, London, UK, pp. 1295-1303,
570 ISBN 978-0-415-48507-4.
571 Davolio, S., Silvestro, F., Gastaldo, T., 2017. Impact of rainfall assimilation on high-
572 resolution hydro-meteorological forecasts over Liguria (Italy). J. Hydrometeor., 18, 2659-
573 2680, doi:10.1175/JHM-D-17-0073.1
574 Dickinson, R., 1988. The force-restore method for surface temperature and its
575 generalization. J. Clim. 1, 1086–1097.
576 Diskin, M.H., Nazimov, N., 1994. Linear reservoir with feedback regulated inlet as a model
577 for the infiltration process. J. Hydrol. 172, 313–330.
578 Gabellani, S., Silvestro, F., Rudari, R., Boni, G., 2008. General calibration methodology for a
579 combined Horton-SCS infiltration scheme in flash flood modeling. Nat. Hazards Earth Sci.
580 8, 1317–1327.
581 Gaume, E., Bain, V., Bernardara, P., Newinger, O., Barbuc, M., Bateman, A., Blaškovicová, L.,
582 Blöschl, G., Borga, M., Dumitrescu, A., Daliakopoulos, I., Garcia, J., Irimescu, A., Kohnova, S.,
583 Koutroulis, A., Marchi, L., Matreata, S., Medina, V., Preciso, E., Sempere-Torres, D., Stancalie,
24 584 G., Szolgay, J., Tsanis, I., Velasco, D., Viglione, A., 2009. A collation of data on European flash
585 floods. J. Hydrol. 367, 70–78. doi: 10.1016/j.jhydrol.2008.12.028.
586 Giannoni, F., Roth, G., Rudari, R., 2000. A semi-distributed rainfall–runoff model based on a
587 geomorphologic approach. Phys. Chem. Earth 25 (7–8), 665–671.
588 Giannoni, F., Roth, G., Rudari, R., 2005. A procedure for drainage network identification
589 from geomorphology and its application to the prediction of the hydrologic response. Adv.
590 Water Resour. 28 (6), 567–581. http://dx.doi.org/ 10.1016/j.advwatres.2004.11.013.
591 Grimaldi, S., Petroselli, A. , Tauro, F., Porfiri, M., 2012. Time of concentration: a paradox in
592 modern hydrology, Hydrol. Sci. J., 57 (2), 217-228, DOI: 10.1080/02626667.2011.644244
593 Guillot P, Duband D. 1967. La méthode du Gradex pour le calcul de la probabilité des crues
594 à partir les pluies. Proceedings Fort Collins Symposium, Vol. 1 of International Hydrology
595 Symposium. Colorado State University: Fort Collins, CO; 506–515.
596 Hapuarachchi, H.A.P., Wang, Q.J., Pagano, T.C., 2011. A review of advances in flash flood
597 forecasting. Hydrol. Processes 25, 2771–2784.
598 Kirpich, Z.P., 1940. Time of concentration of small agricultural watersheds. Civil
599 Engineering, 10 (6), 362.
600 Marchi, L., Borga, M., Preciso, E., Gaume, E., 2010. Characterisation of selected extreme
601 flash floods in Europe and implications for flood risk management. J. Hydrol. 394 (1–2),
602 118–133. doi: 10.1016/j.jhydrol.2010.07.017.
603 MO , 199 . orma . - , Drena e su er icial nstruccio n de carreteras. Direccio n
604 eneral de arreteras, Ministerio de Obras u blicas rbanismo, Madrid.
605 O'Callaghan, J.F., Mark, D.M., 1984. The extraction of drainage networks from digital
606 elevation data. Comput. Vision Graphics Image Process., 28, 328-344.
25 607 Panziera, L., Gabella, M., Zanini, S., Hering, A., Germann, U., Berne, A., 2016. A radar-based
608 regional extreme rainfall analysis to derive the thresholds for a novel automatic alert
609 system in Switzerland. Hydrol. Earth Syst. Sci., 20, 2317–2332.
610 Pavlovic, S., erica, ., t. Laurent, M., Me a, A., 2016. Intercomparison of selected fixed-
611 area areal reduction factor methods. J. Hydrol. 537 (2016) 419–430.
612 doi:10.1016/j.jhydrol.2016.03.027
613 Schaefer, J.T., 1990. The Critical Success Index as an indicator of warning skill. Weather
614 Forecasting 5, 570-575.
615 Silvestro, F., Rebora, N., Ferraris, L., 2009. An algorithm for real-time rainfall rate
616 estimation by using polarimetric radar: RIME. J. Hydrometeorol. 10, 227–240.
617 Silvestro, F., Rebora, N., Ferraris, L., 2011. Quantitative flood forecasting on small and
618 medium size basins: a probabilistic approach for operational purposes. J. Hydrometeorol.,
619 12 (6), 1432-1446.
620 Silvestro, F., Rebora, N., 2012. Operational verification of a framework for the probabilistic
621 nowcasting of river discharge in small and medium size basins. Nat. Hazards Earth Syst.
622 Sci., 12, 763–776.
623 Silvestro, F., Gabellani, S., Delogu, F., Rudari, R., Boni, G., 2013. Exploiting remote sensing
624 land surface temperature in distributed hydrological modelling: the example of the
625 Continuum model. Hydrol. Earth Syst. Sci. 17, 39–62. doi:10.5194/hess-17-39-2013.
626 Silvestro, F., Gabellani, S., Delogu, F., Rudari, R., Laiolo, P., Boni, G., 2015. Uncertainty
627 reduction and parameter estimation of a distributed hydrological model with ground and
628 remote-sensing data. Hydrol. Earth Syst. Sci. 19, 1727–1751. doi:10.5194/hess-19-1727-
629 2015.
26 630 Silvestro, F., Rebora, N.,Giannoni, F., Cavallo, A., Ferraris, L., 2016. The flash flood of the
631 Bisagno reek on 9th October 14 An ‘‘unfortunate” combination of s atial and
632 temporal scales. J. Hydrol. 541, 50–62. doi:10.1016/j.jhydrol.2015.08.004
633 Todini, E., Ciarapica, L., 2001. The TOPKAPI model. In: Mathematical Models of Large
634 Watershed Hydrology (Chapter 12). Water Resources Publications. Singh, V.P. et al. (Eds.),
635 Littleton, Colorado.
636 Versini, P.A., Berenguer, M., Corral, C., Sempere-Torres, D., 2014. An operational flood
637 warning system for poorly gauged basins: demonstration in the Guadalhorce basin
638 (Spain), Nat. Hazards, 71 (3), 1355–1378. doi:10.1007/s11069-013-0949-7.
639 Wooding, R.A., 1965. A hydraulic modeling of the catchment-stream problem. 1. Kinematic
640 wave theory. J. Hydrol., 3, 254–267.
641
642
27 643
644 Table 1. Critical Success Index obtained for the peak values over return period higher than 645 2, 10 and 20 years. For each event, the points of the drainage network used for the 646 computation correspond to the different classification: whole area; drainage area; 647 catchment pertaining; and RH (initial soil moisture).
Event Classification CSI CSI CSI # points T>2 years T>10 years T>20 years (u>0.37) (u>2.25) (u>2.97) total All points (A > 4 km2) 0.541 0.704 0.559 6277 A = 4 – 50 km2 0.583 0.659 0.482 4645 A > 50 km2 0.462 0.791 0.677 1632 Bisagno 1.000 0.982 0.893 56 Entella 0.738 0.630 0.488 231 09/10/2014 Orba - - - 87 Polcevera 0.375 - - 91 Scrivia 0.771 0.931 0.694 104 RH = 0.00 - 0.50 0.087 - - 4791 RH = 0.50 - 0.70 0.131 0.000 0.000 685 RH = 0.70 - 0.90 0.498 0.441 0.373 478 RH = 0.90 - 1.00 0.978 0.814 0.609 323 All points (A > 4 km2) 0.183 0.056 0.156 6277 A = 4 – 50 km2 0.208 0.000 0.000 4645 A > 50 km2 0.118 0.128 0.625 1632 Bisagno 0.000 - 56 Entella 0.448 0.077 0.179 231 10/11/2014 Orba 0.000 - - 87 Polcevera 0.112 0.000 - 91 Scrivia 0.000 - - 104 RH = 0.00 - 0.50 0.000 - - 170 RH = 0.50 - 0.70 0.208 0.000 - 2826 RH = 0.70 - 0.90 0.179 0.070 0.167 3159 RH = 0.90 - 1.00 0.155 0.000 0.000 122 All points (A > 4 km2) 0.454 0.384 0.456 6277 A = 4 – 50 km2 0.471 0.252 0.373 4645 A > 50 km2 0.382 0.971 0.706 1632 Bisagno 0.625 - - 56 Entella 0.000 - - 231 15/11/2014 Orba 0.874 0.447 - 87 Polcevera 0.846 0.667 0.554 91 Scrivia 0.139 0.000 0.000 104 RH = 0.00 - 0.50 0.000 - - 156 RH = 0.50 - 0.70 0.000 0.000 0.000 3830 RH = 0.70 - 0.90 0.357 0.471 0.442 1826 RH = 0.90 - 1.00 0.947 0.324 0.500 465 648
28 649
650 List of figure captions
651 Figure 1. Map of the study area: a) Map from the complete DEM covering the Liguria
652 region and used to apply the warning systems. b) Sub-area where the analysis is focused
653 showing the location of the control points.
654 Figure 2. Point radar-raingauge comparison for 48-h accumulations for the three events:
655 a) Event of 09 October: from 09/10/2014 00:00 to 11/10/2014 00:00; b) Event of 10
656 November: from 10/11/2014 00:00 to 12/11/2014 00:00; c) Event of 15 November: from
657 14/11/2014 00:00 to 16/11/2014 00:00. Regression line and its slope m are included.
658 Figure 3. a) Rainfall accumulation for the event of 09 October from 09/10/2014 00:00 to
659 11/10/2014 00:00 (48-hour accumulation); b) Maximum return period obtained by
660 RFWS; c) Maximum return period obtained by FFWS. Thicker circles highlight the control
661 points selected to show the temporal evolution in subsequent figures.
662 Figure 4. Comparison between RFWS and FFWS outputs for the event of 09 October 2014
663 at Genova control point (Bisagno creek). Warning level from observed flow is also
664 included (threshold discharge values in the right axis). Upstream catchment rainfall is
665 included.
666 Figure 5. Comparison between RFWS and FFWS outputs for the event of 09 October 2014
667 at Panesi control point (Entella river). Warning level from observed flow is also included
668 (threshold discharge values in the right axis). Upstream catchment rainfall is included.
669 Figure 6. Comparison between RFWS and FFWS outputs for the event of 09 October 2014
670 at Montoggio control point (Scrivia river). Upstream catchment rainfall is included.
671 Figure 7. Rainfall accumulation for the event of 10 November from 10/11/2014 00:00 to
672 12/11/2014 00:00 (48-hour accumulation); b) Maximum return period obtained by
29 673 RFWS; c) Maximum return period obtained by FFWS. Thicker circles highlight the control
674 points selected to show the temporal evolution in subsequent figures.
675 Figure 8. Comparison between RFWS and FFWS outputs for the event of 10 November
676 2014 at Panesi control point (Entella river). Warning level from observed flow is also
677 included (threshold discharge values in the right axis). Upstream catchment rainfall is
678 included.
679 Figure 9. Comparison between RFWS and FFWS outputs for the event of 10 November
680 2014 at Carasco control point (Sturla river). Upstream catchment rainfall is included.
681 Figure 10. Rainfall accumulation for the event of 15 November from 14/10/2014 00:00 to
682 16/10/2014 00:00 (48-hour accumulation); b) Maximum return period obtained by
683 RFWS; c) Maximum return period obtained by FFWS. Thicker circles highlight the control
684 points selected to show the temporal evolution in subsequent figures.
685 Figure 11. Comparison between RFWS and FFWS outputs for the event of 15 November
686 2014 at Pontedecimo control point (Polcevera river). Upstream catchment rainfall is
687 included.
688 Figure 12. Comparison between RFWS and FFWS outputs for the event of 15 November
689 2014 at Tiglieto control point (Orba river). Warning level from observed flow is also
690 included (threshold discharge values in the right axis). Upstream catchment rainfall is
691 included.
692 Figure 13. Scatter plot comparison between event maximum outputs from RFWS and
693 FFWS at pixel scale for the 09 October 2014 event. Original return period values have been
694 transformed to the Gumbel reduced variable (double logarithm). a) Grouping is defined by
695 point catchment surface; b) Grouping is defined by catchment pertaining; c) Grouping is
696 defined by upstream catchment soil moisture (relative humidity).
30 697 Figure 14. Scatter plot comparison between event maximum outputs from RFWS and
698 FFWS at pixel scale for the 10 November 2014 event. Original return period values have
699 been transformed to the Gumbel reduced variable (double logarithm). a) Grouping is
700 defined by point catchment surface; b) Grouping is defined by catchment pertaining; c)
701 Grouping is defined by upstream catchment soil moisture (relative humidity).
702 Figure 15. Scatter plot comparison between event maximum outputs from RFWS and
703 FFWS at pixel scale for the 15 November 2014 event. Original return period values have
704 been transformed to the Gumbel reduced variable (double logarithm). a) Grouping is
705 defined by point catchment surface; b) Grouping is defined by catchment pertaining; c)
706 Grouping is defined by upstream catchment soil moisture (relative humidity).
707 Figure 16. Same figure as Fig. 4 (control point of Genova in Bisagno creek), including the
708 additional outputs from RFWS using the rainfall duration related to an alternative
709 formulation for the concentration time.
710
711
31 Figure1
44.4 Genova
Settepani radar Savona 44.2
lat (º) lat La Spezia (a) 44.0 Massa 20km 20 km
43.8 Sanremo 8.0 8.5 9.0 9.5 10.0 long (º) Tiglieto Montoggio Carasco Orba river Scrivia river Sturla river (71.4 km 2) (59.4 km 2) (131.6 km 2)
Pi ota
Stura di Ovada
44.5
Lavagna
lat (º) lat 44.4 Pontedecimo Polcevera creek (60.9 km 2)
10km Genova Bisagno creek 44.3 (b) (90.9 km 2) 10 km Panesi 8.6 8.8 9.0 9.2 Entella river 9.4 long (º) (370.6 km 2) Figure2
201410090000-201410110000 201411100000-201411120000 201411140000-201411160000 250 400 m = 0.98 m = 0.69 m = 0.65 250 200
300 200 150 150 200 100 Radar(mm) Radar(mm) Radar(mm) 100
100 50 50 (a) (b) (c) 0 0 0 0 100 200 300 400 0 50 100 150 200 250 0 50 100 150 200 250 Raingauge (mm) Raingauge (mm) Raingauge (mm) Figure3 Accumulated Rainfall
44.4
44.2 lat (º) lat
44.0
0 60 120 180 240 300 360 420 480 540 600 43.8 (a) P (mm) 8.0 8.5 9.0 9.5 10.0 long (º) Warning Level RFWS
44.5
lat (º) lat 44.4
5 10 30 50 100 200 500 44.3 (b) T (years) 8.6 8.8 9.0 9.2 9.4 long (º) Warning Level FFWS
44.5
lat (º) lat 44.4
5 10 30 50 100 200 500 44.3 (c) T (years) 8.6 8.8 9.0 9.2 9.4 long (º) Figure4
Warning Level: Genova-Bisagno 0 20 2 RFWS 40 500 A = 90.9 km FFWS 60 80 P(mm/h) 200 Qobs 100
100 1027
50 853
30 725 Q (m3/s) Q Warninglevel(years) T 10 450
5 275
0 0 16: 17: 18: 19: 20: 21: 22: 23: 00: 01: 02: 03: 04: 05: 06: 09 10 October 2014 Figure5
Warning Level: Panesi-Entella 0 20 RFWS 40 500 FFWS 60 80 P(mm/h) 200 Qobs 100 A = 370.6 km2 100
50 2186
30 1158 Q (m3/s) Q Warninglevel(years) T 10 1152
5 704
0 0 16: 17: 18: 19: 20: 21: 22: 23: 00: 01: 02: 03: 04: 05: 06: 09 10 October 2014 Figure6
Warning Level: Montoggio-Scrivia 0 20 RFWS (years) 40 500 FFWS (years) 60 80 P(mm/h) 200 A = 59.4 km2 100
100
50
30 Warninglevel(years) T 10
5
0 16: 17: 18: 19: 20: 21: 22: 23: 00: 01: 02: 03: 04: 05: 06: 09 10 October 2014 Figure7 Accumulated Rainfall
44.4
44.2
lat (º) lat (a) 44.0
0 60 120 180 240 300 360 420 480 540 600 43.8 P (mm) 8.0 8.5 9.0 9.5 10.0 long (º) Warning Level RFWS
44.5
lat (º) lat 44.4
5 10 30 50 100 200 500 44.3 (b) T (years) 8.6 8.8 9.0 9.2 9.4 long (º) Warning Level FFWS
44.5
lat (º) lat 44.4
5 10 30 50 100 200 500 44.3 (c) T (years) 8.6 8.8 9.0 9.2 9.4 long (º) Figure8
Warning Level: Panesi-Entella 0 20 2 RFWS 40 500 A = 370.6 km FFWS 60 80 P(mm/h) 200 Qobs 100
100
50 2186
30 1158 Q (m3/s) Q Warninglevel(years) T 10 1152
5 704
0 0 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 00: 01: 02: 03: 04: 10 11 November 2014 Figure9
Warning Level: Carasco-Sturla 0 20 2 RFWS (years) 40 500 A = 131.6 km FFWS (years) 60 80 P(mm/h) 200 100
100
50
30 Warninglevel(years) T 10
5
0 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 00: 01: 02: 03: 04: 10 11 November 2014 Figure10 Accumulated Rainfall
44.4
44.2 lat (º) lat (a) 44.0
0 60 120 180 240 300 360 420 480 540 600 43.8 P (mm) 8.0 8.5 9.0 9.5 10.0 long (º) Warning Level RFWS
44.5
lat (º) lat 44.4
5 10 30 50 100 200 500 44.3 (b) T (years) 8.6 8.8 9.0 9.2 9.4 long (º) Warning Level FFWS
44.5
lat (º) lat 44.4
5 10 30 50 100 200 500 44.3 (c) T (years) 8.6 8.8 9.0 9.2 9.4 long (º) Figure11
Warning Level: Pontedecimo-Polcevera 0 20 RFWS (years) 40 500 FFWS (years) 60 80 P(mm/h) 200 A = 60.9 km2 100
100
50
30 Warninglevel(years) T 10
5
0 02: 03: 04: 05: 06: 07: 08: 09: 10: 11: 12: 13: 14: 15: 16: 15 November 2014 Figure12
Warning Level: Tiglieto-Orba 0 20 RFWS 40 500 A = 71.4 km2 FFWS 60 80 P(mm/h) 200 Qobs 100
100 1
50 1
30 1 Q (m3/s) Q Warninglevel(years) T 10 377
5 231
0 0 02: 03: 04: 05: 06: 07: 08: 09: 10: 11: 12: 13: 14: 15: 16: 15 November 2014 Figure13
Event 20141009 Surface grouping Event 20141009 Catchment grouping 7 7 area (km2) 4-50 n points:4645 Bisagno n points:56 n points total:6277 area (km2) >50 n points:1632 Scrivia n points:104 6 6 Entella n points:231 Polcevera n points:91 Orba n points:87 5 5
4 4
RFWS u (-) u RFWS 3 (-) u RFWS 3
2 2
1 1 (a) (b) 0 T=3.2 T=7.9 T=20.6 T=55.1 T=148 T=404 0 T=3.2 T=7.9 T=20.6 T=55.1 T=148 T=404 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 FFWS u (-) FFWS u (-)
Event 20141009 RH grouping 7 RH (%) 0-50 npoints:4791 RH (%) 50-70 npoints:685 6 RH (%) 70-90 npoints:478 RH (%) 90-100 npoints:323
5
4
RFWS u (-) u RFWS 3
2
1 (c) 0 T=3.2 T=7.9 T=20.6 T=55.1 T=148 T=404 0 1 2 3 4 5 6 7 FFWS u (-) Figure14
Event 20141110 Surface grouping Event 20141110 Catchment grouping 7 7 area (km2) 4-50 n points:4645 Bisagno n points:56 n points total:6277 area (km2) >50 n points:1632 Scrivia n points:104 6 6 Entella n points:231 Polcevera n points:91 Orba n points:87 5 5
4 4
RFWS u (-) u RFWS 3 (-) u RFWS 3
2 2
1 1 (a) (b) 0 T=3.2 T=7.9 T=20.6 T=55.1 T=148 T=404 0 T=3.2 T=7.9 T=20.6 T=55.1 T=148 T=404 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 FFWS u (-) FFWS u (-)
Event 20141110 RH grouping 7 RH (%) 0-50 npoints:170 RH (%) 50-70 npoints:2826 6 RH (%) 70-90 npoints:3159 RH (%) 90-100 npoints:122
5
4
RFWS u (-) u RFWS 3
2
1 (c) 0 T=3.2 T=7.9 T=20.6 T=55.1 T=148 T=404 0 1 2 3 4 5 6 7 FFWS u (-) Figure15
Event 20141115 Surface grouping Event 20141115 Catchment grouping 7 7 area (km2) 4-50 n points:4645 Bisagno n points:56 n points total:6277 area (km2) >50 n points:1632 Scrivia n points:104 6 6 Entella n points:231 Polcevera n points:91 Orba n points:87 5 5
4 4
RFWS u (-) u RFWS 3 (-) u RFWS 3
2 2
1 1 (a) (b) 0 T=3.2 T=7.9 T=20.6 T=55.1 T=148 T=404 0 T=3.2 T=7.9 T=20.6 T=55.1 T=148 T=404 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 FFWS u (-) FFWS u (-)
Event 20141115 RH grouping 7 RH (%) 0-50 npoints:156 RH (%) 50-70 npoints:3830 6 RH (%) 70-90 npoints:1826 RH (%) 90-100 npoints:465
5
4
RFWS u (-) u RFWS 3
2
1 (c) 0 T=3.2 T=7.9 T=20.6 T=55.1 T=148 T=404 0 1 2 3 4 5 6 7 FFWS u (-) Figure16
Warning Level: Genova-Bisagno 0 RFWS using D = 3 h (Tc Kirpich = 2.5 h) 20 2 40 500 A = 90.9 km RFWS using D = 6 h (Tc Témez = 5.9 h) 60 FFWS 80 P(mm/h) 200 100 Qobs 100 1027
50 853
30 725 Q (m3/s) Q Warninglevel(years) T 10 450
5 275
0 0 16: 17: 18: 19: 20: 21: 22: 23: 00: 01: 02: 03: 04: 05: 06: 09 10 October 2014