The RainBO Platform for Enhancing Urban Resilience to Floods: An E ﬃ cient Tool for Planning and Emergency Phases

: Many urban areas face an increasing ﬂood risk, which includes the risk of ﬂash ﬂoods. Increasing extreme precipitation events will likely lead to greater human and economic losses unless reliable and e ﬃ cient early warning systems (EWS) along with other adaptation actions are put in place in urban areas. The challenge is in the integration and analysis in time and space of the environmental, meteorological, and territorial data from multiple sources needed to build up EWS able to provide e ﬃ cient contribution to increase the resilience of vulnerable and exposed urban communities to ﬂooding. E ﬃ cient EWS contribute to the preparedness phase of the disaster cycle but could also be relevant in the planning of the emergency phase. The RainBO Life project addressed this matter, focusing on the improvement of knowledge, methods, and tools for the monitoring and forecast of extreme precipitation events and the assessment of the associated ﬂood risk for small and medium watercourses in urban areas. To put this into practice, RainBO developed a webGIS platform, which contributes to the “planning” of the management of river ﬂood events through the use of detailed data and ﬂood risk / vulnerability maps, and the “event management” with real-time monitoring / forecast of the events through the collection of observed data from real sensors, estimated / forecasted data from hydrologic models as well as qualitative data collected through a crowdsourcing app.


Introduction
Climate change affects the water cycle by intensifying it and this can change the magnitude, frequency, and timing of river floods in areas of the planet [1] such as some parts of Europe [2]. In particular, Blöschl et al. [3,4] by analyzing a pan-European database over the past five decades found clear patterns of change in flood timing in Europe due to changes in climate. However, the increasing trend of disasters due to floods in Europe is due also to non-climatic drivers, such as continued socio-economic growth, which induces population growth, economic wealth, and unplanned The structure of the article is as follows: Section 2 describes an examination of existing data and useful models for the purposes of the RainBO platform, and afterward, the innovative methodologies developed within the project are explained. In Section 3, the platform and the main results achieved during the project are then presented, with a description of its main structure, databases, and modules. Finally, remarks for further developments are described.

Study Areas
The application of the RainBO platform is performed on two test areas, the cities of Bologna and Parma, located in the Emilia-Romagna region, Italy. The climatic features referred to 1991-2015 period for these two cities are shown in Table 1 [25].    As shown in the figure, the catchment of the Ravone creek is small, whereas the catchment of the Parma river is medium-large. The different sizes of the catchments are reflected in two different modeling approaches of the hydrologic forecasts, as explained below.
It is worthy to mention that each hydraulic section of the Emilia-Romagna catchments has three specific thresholds/levels of alarm: yellow threshold/warning, orange threshold/pre-alarm and red threshold/alarm. These thresholds are site-specific and defined according to Civil Protection purposes for public safety, taking into account the geometry of the hydraulic section of the watercourse and the statistical distribution of historical recordings.

Ravone Catchment
The Ravone catchment, together with the Aposa catchment, is the largest river basins at SW of Bologna. They flow from the hills directly across the central part of the city. The upper part of the Ravone catchment is characterized by hills and steep slopes, with a prevalent vegetation coverage (grass, shrubs, and forest) and partial urbanization in the stream valley. In contrast, the lower portion is flat and densely urbanized, the drainage network is mainly artificial, and the watercourse is connected with the main urban drainage system in a critical spillway crosspoint, flowing in a culvert underneath the city's urban area before joining the Reno river. The length of the natural reach of the Ravone is approximately 4 km covering an area of 6 km 2 .
The Ravone catchment has been chosen as a study area of RainBO because, as recorded in historical data, the response of the catchment to extreme rainfall past events caused significant damages. For instance, during a flood that occurred on July 22nd, 1932, a victim and severe damages to streets and houses in the Southern part of the city were recorded [26].
The catchment is equipped by an existing monitoring network (Figure 2), including a weather station 500 m far from the Ravone catchment equipped with a rain gauge on S. Luca hill where data have been collected since the early 1930s. To integrate the available dataset and to better monitor the Ravone catchment, in 2014 a second rain gauge was installed at the catchment upstream end of mount Paderno and a water level gauge was installed at the culvert entry of the Ravone creek ( Figure 3). It is worthy to mention that each hydraulic section of the Emilia-Romagna catchments has three 169 specific thresholds/levels of alarm: yellow threshold/warning, orange threshold/pre-alarm and red

182
The Ravone catchment has been chosen as a study area of RainBO because, as recorded in 183 historical data, the response of the catchment to extreme rainfall past events caused significant 184 damages. For instance, during a flood that occurred on July 22 nd , 1932, a victim and severe damages 185 to streets and houses in the Southern part of the city were recorded [26].

209
The available monitoring network for the Parma basin is composed of 15 rain gauges, 11 210 thermometers, and 6 water level gauges, these sensors are sufficient for the application of the RF
212 Figure 3. The entry of the culvert of the Ravone creek where the water level gauge is installed and the alarm thresholds are highlighted.

Parma Catchment
Parma river is the main river flowing through a homonymous city in Emilia-Romagna; it starts from Mount Marmagna at 1842 m.a.s.l. and flows in an N-NE direction joining the River Po near the city of Colorno as the right tributary. Usually, the hydrological responses for these basins are characterized by high discharges in the spring and autumn and low discharges in the summer.
The small catchment area and the steep part of the valley give high hydrological response during severe storms, generating, under particular conditions, flash flood events. Its major tributary is the Baganza River which has a very similar course and joins the main river on its left in the city of Parma ( Figure 4).

209
The available monitoring network for the Parma basin is composed of 15 rain gauges, 11 210 thermometers, and 6 water level gauges, these sensors are sufficient for the application of the RF

211
Model, hence, no other installation was required ( Figure 5). The available monitoring network for the Parma basin is composed of 15 rain gauges, 11 thermometers, and 6 water level gauges, these sensors are sufficient for the application of the RF Model, hence, no other installation was required ( Figure 5).

230
the main legislative reference is the Floods Directive [6]. In the directive, the hydraulic risk is the 231 product of the hazard and potential damage at a specific event: 232 where:

233
• P (Hazard): it is the probability of occurrence, within a certain area and in a certain

Background of the RainBO Project
The existing data and models at the beginning of the project identified as necessary for the development of the RainBO platform are described.

Territorial Data
According to the existing national and regional guidelines for the planning of Civil Protection, the following maps on the Emilia-Romagna region are available: These data are projected in the WGS84 UTM32N reference system.

Hazard, Vulnerability, and Risk Maps
To set up a tool for the management of flood risk in the Member States of the European Union, the main legislative reference is the Floods Directive [6]. In the directive, the hydraulic risk is the product of the hazard and potential damage at a specific event: where: • P (Hazard): it is the probability of occurrence, within a certain area and in a certain time interval, of a natural phenomenon of assigned intensity • E (Exposure): it represents people and/or assets (structures, infrastructures, etc.) and/or activities (economic, social, etc.) exposed to a natural event • V (vulnerability): the degree of capacity (or incapacity) of a system/element to resist at the natural event • Dp (potential damage): it is considered as the degree of foreseeable loss following a natural phenomenon of a given intensity, the function of both value and vulnerability of the exposure • R (risk): expected number of victims, injured persons, damage to property, cultural assets e environmental, destruction or interruption of economic activities, as a result of a natural phenomenon of assigned intensity Emilia-Romagna Region has developed the Flood Risk Management Plan (FRMP), to be compliant with the Floods Directive [6] and Legislative Decree 49/2010, which requires these flood risk management plans to include measures to reduce the probability of flooding and its potential consequences and to address all phases of the flood risk management cycle but in particular the prevention, the protection, and the preparedness. As the causes and consequences of floods are different in the different member states of the Community, the Management Plans take into account the specific characteristics of the territories and propose specific objectives and measures tailored to the needs and priorities.
The FRMP of the Emilia-Romagna Region is represented by three projects, one for each hydrographic district (Po River, Northern Apennines, and Central Apennines).
Existing hazard maps represent the potential extent of flooding caused by watercourses (natural and artificial) with reference to three scenarios (rare floods, infrequent, and frequent) colored with three different shades intensity of blue, depending on the frequency of flooding as follows: • rare floods of extreme intensity: return time up to 500 years from the event (low probability) • infrequent floods: return time between 100 and 200 years (average probability) • frequent floods: return time between 20 and 50 years (high probability) The hazard maps from the Floods Directive are available on the whole Italian territory. They constitute the reference hazard maps for the computation of hydraulic risk and, for the purpose of the RainBO project, they have been selected on the study areas of Bologna and Parma.
Moreover, for the Bologna study case, in addition to the reference hazard map where the Ravone and other small streams are not included, a specific map for the Ravone creek is available [27]. This has been obtained through scenarios of flood analysis [28].
The existing vulnerability maps from the Floods Directive have been clipped on the study areas of Bologna and Parma. It should be noted that these reference maps do not consider in detail population distribution issues, as well as risk maps, as a consequence. In particular, to define the expected damages of a flood, the Directive suggests including the following main items: The risk maps indicate the presence of potentially exposed elements (population involved, services, infrastructure, economic activities, etc.) which fall within floodable areas by means of a classification in 4 risk categories, represented by a color scale: yellow (moderate or no risk), orange (medium risk), red (high risk), purple (very high risk).
Existing risk maps from Floods Directive, created by the integration of hazard maps and vulnerability ones, identify static situations and do not take into account urban territory resilience peculiarity. In this case, risk maps have also been selected for the study areas of Bologna and Parma.

Historical Events
A catalog of historical events from 1981 is available for the Parma and Reno, including the Ravone catchment. A historical event is defined as the exceedance of at least one pre-alarm threshold. For each event, the involved catchment, the exceeded thresholds of water level gauge, as well as the description of the flooded area are recorded.
Moreover, where available, for each event further information (i.e., the number of people evacuated, dead or wounded people, emergency state-if requested, the possible assessment of economic damages) were collected. Moreover, the reports for all the events that occurred in Emilia-Romagna are available [29].

Observed Meteorological Data
A complex infrastructure for environmental monitoring and for an early-warning system is based on the integration of different data from many sources. The hydro-pluviometric network of Arpae collects data from rain gauges, thermometers, and water level gauges on the Emilia-Romagna region ( Figure 6).      thousands of sensors distributed on the territory and builds a digital map over time [30].

311
The development of the RainBO platform implied also the empowerment of the monitoring For the RainBO purposes, the key variables are water level gauges (242 sensors) and rain gauges (282 sensors). Data coming from the hydro-pluviometric monitoring system are available on the Arpae ftp server. These data are acquired from the network every 15 min, the units are millimeters (mm) for precipitation and meters (m) for water level. In general terms, the series starts from 2003, the historical series for some sensors start from 1980.
These data are integrated into the Sensornet platform through a web service. Sensornet is the Internet of Things Platform of the Emilia-Romagna Region, that collects data and information from thousands of sensors distributed on the territory and builds a digital map over time [30].
The development of the RainBO platform implied also the empowerment of the monitoring network in critical areas not properly covered by sensors such as the Ravone catchment. For this reason, new monitoring sensors were installed in the Ravone area: • a new real-time water level gauge has been installed in the upper part of the river basin • a new weighing rain gauge has been located in a public property on the right side of the valley These new monitoring sensors installed in the Ravone area are aimed at measuring environmental variables to be collected in the RainBO platform and at monitoring the Ravone creek with a high degree of accuracy.

Forecast Meteorological Data
The COSMO-LAMI forecasts provided by Arpae-SIMC are available on an open data platform as GRIB files over Emilia-Romagna [31].
Meteorological COSMO-LAMI forecasts are based on the operative non-hydrostatic limited-area atmospheric COSMO (Consortium for Small-scale Modelling) model, nested on the ECMWF (European Centre for Medium-Range Weather Forecasts) operational global forecast. This system is developed and maintained by the homonymous European consortium, and managed by Arpae-SIMC on the basis of the LAMI (Limited Area Model Italia) agreement. The forecast is issued twice a day (00 and 12 UTC) with a time range of 72 h on a regular 5 × 5 km grid covering the Mediterranean area.

Estimated Data
Commercial microwave links and radar data are estimated data, essential for the purposes of RainBO in order to monitor precipitation events.
The microwave links based on the rainfall monitoring system is an innovative but already tested technology, exploiting the microwave links used in commercial cellular communication networks (so-called commercial microwave links, CML).
Conventional rain gauges are not so effective during intense precipitation as their operating principle is based on mechanical tilting parts, which makes their measurements unreliable during this type of phenomenon. Rain gauges provide point-like measurements of the amount of rain fallen within the instrument sampling area, cumulated on time intervals, usually ranging from one minute to one day, with well known instrumental [32] and representativeness [33] limitations.
A relatively new and independent approach to the estimates of precipitation at the ground became available in the last decades with the broad diffusion of CMLs for cellular communication: integral precipitation content along a line path between two antennas can be estimated by measuring the attenuation of the microwave signal along the same path [34].
Heavy rain causes electromagnetic signal attenuation (from the transmitting antenna to the receiving one) and, subsequently, path-averaged rainfall intensity can be retrieved from the signal's attenuation between transmitter and receiver by applying, almost in real-time, a rainfall retrieval algorithm. A distributed monitoring system can be developed by using received signal level data from the massive number of CMLs used worldwide in commercial cellular communication networks.
The first studies on this technology were concentrated on algorithms for spatial-temporal interpolation [35] from the joint analysis of multiple CMLs. The great potential of CMLs for ungauged regions was demonstrated by the Burkina Faso application [36]. In 2012, Overeem [37] demonstrated that processing algorithms are capable of providing real-time rainfall maps for an entire country, in this case, the Netherlands.
With regard to radar data, an existing dataset on Emilia-Romagna is based on hourly precipitation estimates obtained from the merger of the regional radar network managed by Arpae-SIMC. This network is composed of two C-Band systems, one located at San Pietro Capofiume (Bologna) and the other in Gattatico (Reggio Emilia). Every 5 min during precipitation events, the radars provide reflectivity data that are processed by several algorithms. The reflectivity value is correlated to the precipitation intensity. Reflectivity data are provided by Arpae-SIMC in open data [38].
For the RainBO project, an ad hoc stream was triggered for the automatic and periodic distribution of both reflectivity maps and hourly precipitation maps on the ftp server provided by Lepida, according to the RainBO project goals. Sent images are a merge of the two systems (reflectivity radar maps and accumulated hourly precipitation). Every new capture of the reflectivity (the frequency of observations is related to the weather conditions at the time) generates a merged image, which is then sent to the Lepida ftp server.

Crowdsourcing
A platform addressed to provide information for planning and management of extreme events and flash floods in urban areas should comprise also observations and contributions from citizens. These types of data, usually collected by means of applications, are a source of additional information during ongoing events but they are also conceived in order to engage and raise awareness of citizens. In this regard, the system Rmap [39] is a participatory monitoring and exchange system promoted by Arpae-SIMC, based on open hardware and software infrastructures, to collect and share meteorological data gathered by citizens between public and private institutions.
The Rmap application is mainly addressed to users and citizens with meteorological domain expertise and collects automatically the weather data in a WMO (World Meteorological Organization) binary data software (Binary Universal Form for the Representation of meteorological data-BUFR) through dedicated devices based on open hardware and free software. The weather information data can also be uploaded manually by expert users by using an on-line application, but the graphical user interface is not conceived as a smart tool for the general public.
It defines a set of standards for meteorological data sensing (security, reliability, elaboration) and for the transmission data system (transmission protocols, data formats, metadata formats, etc.).
The Rmap project has been promoted by Arpae-SIMC for some years, as it is an interesting project with the objective of defining methods, protocols, and formats to collect and share environmental data. The project is also promoted by Arpa Veneto, Cineca and the Computer Science Department of the University of Bologna and the RaspiBO network.
The Rmap system was taken into consideration because of this robust partnership and the relevant effort spent in its standardization.
The Rmap project adopts indeed a scientific approach based on standards defined by the WMO, in particular using their elaboration and classification process.

Models
The models used for the development of the RainBO services are: [40,41] is a one-dimensional model developed by Arpae simulating the soil water balance, nitrogen balance, and crop development. The model is usually applied to agricultural case studies, nonetheless one of its main outputs (i.e., soil moisture) is a crucial variable for hydrological purposes. The CRITERIA-1D model simulates soil water movement by using a simplified model (tipping bucket) or a numerical model. It requires as an input at least daily data of temperature and precipitation, soil features, and crop information.
CRITERIA-3D [42] is a physically-based model developed by Arpae that works at the catchment-scale and solves equations of surface and subsurface water flow in a three-dimensional domain. The hydrologic component is a dynamic link library integrated into the other Arpae software, implemented within a comprehensive model that simulates the physical processes occurring in the catchment: surface energy, radiation budget, snow accumulation and melt, potential evapotranspiration, plant development, and plant water uptake.
The two models are under development and they are available as open-source code [43]. After the flooding of the Parma and Baganza rivers in Parma on October 2014, caused by heavy rains, the Civil Protection Agency of the Emilia-Romagna region required Arpae to develop a hydrological simulation model capable of promptly evaluating in advance the probability of overcoming the alert thresholds, especially for rapid or flash flood events. The RANDOM FOREST (RF) algorithm, applied in a hydrological context, provides the probability of overcoming the alert thresholds of some observation points for basins at small and medium scales in the oncoming next 6-8 h.
The RF model was added beside the existent hydrological-hydraulic model applied in real-time into the Flood Early Warning System Emilia-Romagna (FEWS-EMR) to provide a fast and preliminary response during flash flood or extreme rainfall events in the Emilia-Romagna basins.
The model is an ensemble learning method that operates by constructing a multitude of decision trees at training time and outputting the mode of the classes (classification) or mean prediction (regression) of the individual trees. Each tree classifies the dataset using a subset of variables. The number of trees in the forest and the number of variables in the subset are hyper-parameters and, for this reason, they have to be chosen a priori.
The number of trees is in the order of hundreds, while the subset of variables is quite small, if compared to the total number of variables, in Figure 7.the final Random Forests tree generated for Parma River at Ponte Verdi is shown, all paths end with terminal node that contains the probability of exceedance for each H.T.A. (hydrometric thresholds alert).

434
RF also provides a natural way to assess the importance of input variables (predictors). This is 435 achieved by removing a variable at a time and assessing whether the out-of-bag error changes or not.

436
If the out-of-bag error changes, the variable is important for the decision [44]. RF also provides a natural way to assess the importance of input variables (predictors). This is achieved by removing a variable at a time and assessing whether the out-of-bag error changes or not. If the out-of-bag error changes, the variable is important for the decision [44].
Model parameterization was performed using historical data (2003-2016), while recent data (2017-2019) are used for validation. Model parametrization was accomplished primarily by extracting historical events for each river section, hence defining the reference response time of the basin. This value was then used in the RF model, defining the maximum aggregation time for rainfall ( Figure 8). For the RainBO project, the model was implemented using as input the same observed date used for the other Hydrological-Hydraulic model: observed mean hourly rainfall and discharge data.
The code is an open-source R package available for free download on GitHub [45].

461
The implementation of the algorithm for the specific needs of the Italian test areas, as well as 462 code debugging and improving, was carried out within the RainBO project.

Foreground of RainBO Project: Innovation and Development
The development of the RainBO infrastructure started from existing data and models, then it focused on the application of new models specifically developed and new technologies, such as CML to enhance the monitoring systems during extreme events.
The project includes the development of a crowdsourcing web application for collecting and sharing information on the observed weather and its possible local effects/impact.

Commercial Microwave Links
The rainfall monitoring system proposed in the RainBO project exploits the Commercial Microwave links (CML or simply Microwave links-MWL) used worldwide in commercial cellular communication networks. Rain-induced attenuation and, as a consequence, path-averaged rainfall intensity can be retrieved from the signal's attenuation by applying, almost in real-time, a rainfall retrieval algorithm. The algorithm chosen for this purpose is the RAINLINK retrieval algorithm [37]. The code is an open-source R package available for free download on GitHub [45].
The implementation of the algorithm for the specific needs of the Italian test areas, as well as code debugging and improving, was carried out within the RainBO project.
To validate the algorithm, a dataset was provided by Vodafone Italia on Bologna and Parma urban areas from February 2016 to June 2016. Other (not commercial) microwave link data were supplied by Lepida to cover the Apennines area, from March 2016 up to date, providing near-real-time data.
During the project, the CML data validation on the Bologna area was performed by comparing the quantitative precipitation estimates (QPE) from CML, radar, and rain gauges.
Radar and rain gauges data were chosen for the validation because they are currently validated, used, and published by ARPAE in their operative meteorological services. It resulted that the coherence between CML and the other estimates is quite promising even if it requires a tuning activity to integrate the dataset with existing technologies (like rain gauges or radar).
Excellent results are achieved mainly in a convective event as shown in Figure 9. On 11 May 2016 precipitation occurred with a well-defined gradient in the West-East direction and some local maxima in the North-West and South-West part of the province. Microwave accumulation slightly underestimated the rainfall field while an overestimation is recorded in the non-adjusted radar. The adjustment procedure well calibrates radar data as displayed in the top right panel of Figure 9. Fine-scale structure of the daily amount is well detected in both remote sensing maps.

493
The precipitation monitoring infrastructure implemented during the RainBO project, based on 494 the CML data, was called Rainlink4EMR (as the RAINLINK algorithm was applied to the Emiliaattenuation raw data (CML data) from the Lepida telecommunication infrastructure and provides 497 rainfall estimates on the midpoint of the links with a delay of a few minutes.

498
The validation study allowed us to achieve a first operational version of the related service,

499
providing satisfactory reliability in monitoring convective events. A qualitative analysis showed that the performance of the CLM estimate seems to increase in the second half of the analyzed period (May and June), mainly characterized by convective storms, even embedded in frontal patterns, where showers and heavy rain play an important role. As the microwave estimate is based on the attenuation that occurred in the link path, it is strongly related to rainfall intensity and the signal is, normally, stronger in the convective season. This behavior was confirmed by the quantitative analysis done by using statistical indicators that confirm a good rainfall estimation by CML data during summer and spring seasons. The validation has pointed out that CML estimation slightly underestimates precipitation occurrence both in spatial coverage and point amount, while radar adjusted has a complementary feature.
The precipitation monitoring infrastructure implemented during the RainBO project, based on the CML data, was called Rainlink4EMR (as the RAINLINK algorithm was applied to the Emilia-Romagna region) and a near-real-time module was implemented, which downloads the power attenuation raw data (CML data) from the Lepida telecommunication infrastructure and provides rainfall estimates on the midpoint of the links with a delay of a few minutes.
The validation study allowed us to achieve a first operational version of the related service, providing satisfactory reliability in monitoring convective events.

Crowdsourcing App
A crowdsourcing web application was implemented in RainBO to collect and display information regarding the observed current weather from expert users, as well as from people without any technical skills. This webApp was created by means of a networking activity with the Rmap project [39] and, thus, it was called Rmap4RainBO.
The Rmap application is mainly addressed to users with meteorological domain expertise and collects automatically data coming from standard hardware devices and weather observation manually uploaded. The RainBO crowdsourcing application, called Rmap4RainBO, develops and improves the Rmap functionalities for the uploading and visualization of the observed weather and it keeps the WMO standard in the weather code. The RainBO crowdsourcing component, differently from the Rmap one, aims to address people without any technical skills so, as a consequence, it was developed as an intuitive and smart application that can be accessed through the RainBO project homepage [46].
The Rmap4RainBO application can benefit from the Rmap data and vice versa as the crowdsourcing information uploaded by citizens through Rmap4RainBO feeds the Rmap database.
Differently from the Rmap, the RainBO crowdsourcing webApp records the impacts, meant as effects of weather on the territory (damaged roads, fallen trees, ice on the road, etc.). The codes used to label the impacts were defined internally at the RainBO consortium as any official code was found on WMO standards; it represents a further innovative contribution by the RainBO project.

RainBO Vulnerability Model
The vulnerability module calculates the degree of vulnerability of exposed items over flood events. The vulnerability reference maps do not consider in detail population distribution issues, whereas the vulnerability model developed for RainBO includes the presence of sensible targets in the territory (e.g., schools, nurseries, hospitals) and critical targets that can worsen a scenario reducing resilience, such as the fire brigade building.
To take into account in a realistic way the distribution of people on a territory, the developed algorithm considers different time frames: for example, the distribution of the population is supposed to be different in night hours (mainly in houses) with respect to working hours (mainly in workplaces). In a similar way, during morning time, students and teachers are supposed to be in schools, while during the afternoon school users decrease, and during the night no one is supposed to occupy these target buildings.
Vulnerability maps calculated by the vulnerability module are based on territorial data collected on the platform. Moreover, the maps, in summary, are calculated as a function of: This information is useful not only to compute more realistic vulnerability maps but also because it can be used by the early warning module. Through these data, the priority of targets to be warned during an ongoing emergency can be identified.

Hydrologic Forecast for Small Catchments
One of the main activities of the project has been the development of a model of the hydrological forecast for small basins, by using Ravone as a test case.
Within the available dataset (from 2014 to 2018) of observed water level at the culvert entry there are not events that exceed the alarm threshold. For this reason, CRITERIA-3D model has been used to simulate scenarios in order to assess the effects of severe rainfall events potentially able to exceed the alarm threshold (Figure 10), starting from initial conditions of soil moisture corresponding to the most remarkable event within the dataset (recorded on March 25th, 2015). The scenarios include three possible precipitation sum (70, 85, and 100 mm per event) with two possible event lengths (9 or 14 h) and two possible precipitation hyetographs (triangular and trapezoidal). These choices correspond to a discretization of the precipitation intensities recorded during the most remarkable past events (including the event recorded in 1932) when water level gauges were not installed. Thus, 12 possible precipitation scenarios on the Ravone catchment have been simulated with the CRITERIA-3D model. This discretization is a compromise between the need to simulate as many cases as possible and to run simulations within an acceptable computational time.

567
The water levels simulated by CRITERIA-3D have been integrated with observed data. On the 568 resulting dataset, a statistical analysis has been performed, taking into account water levels, 569 precipitation and soil moisture (defined as water holding capacity, see below for further details). As 570 a result, a significant logistic regression between these variables has been identified.

571
The hydrological forecast is based on this regression using as input the forecast of precipitations   The water levels simulated by CRITERIA-3D have been integrated with observed data. On the resulting dataset, a statistical analysis has been performed, taking into account water levels, precipitation and soil moisture (defined as water holding capacity, see below for further details). As a result, a significant logistic regression between these variables has been identified.
The hydrological forecast is based on this regression using as input the forecast of precipitations of the COSMO-LAMI model, presenting the best available resolution in open data and calibrated on the study area. In order to include the spatial variability of the event, the computation is performed on the Cosmo grid cell containing the prevailing area of the Ravone catchment and on the 8 neighboring cells.
To validate the forecast algorithm, a hindcast analysis using the COSMO-LAMI forecasts for the period 2015-2016 was carried out. The analysis showed that the maximum intensity of precipitation forecast on the area is mainly underestimated, with an average underestimation of approximately a third of the observed value. To compensate for this underestimation, that could be cause missing alarms, the operational dataset used as input for the hydrologic forecast is integrated with a second forecast series where the maximum hourly intensity of precipitation is increased by 33%.
Therefore, an ensemble of forecast scenarios of precipitation is produced and an ensemble of the hydrological forecast is derived. The statistical distribution of this output is computed to provide a boxplot of the hydrological forecast.
In addition to the precipitation, the crucial variable for the hydrological forecasts is the soil moisture of the catchment. We decided to use an estimated value of this variable instead of measured one because it has the advantage that it is not affected by sensor lacking or failures and local peculiarity. To estimate the soil moisture for catchments of small dimensions as Ravone, it is possible to assess the mean soil moisture of the area by means of a mono-dimensional soil water balance model, as CRITERIA-1D. For the development of the algorithm for the forecasts of exceeding the hydrometric threshold, a new output variable named water holding capacity (WHC) has been added to the CRITERIA-1D model. WHC provides the maximum amount of water the soil can retain before the runoff starts, given the current conditions of soil moisture. For the Ravone study case, CRITERIA-1D is set with the parameters of the prevailing soil on the catchment (silty loam) and the parameters of prevailing crop coverage in the area (fallow); the WHC index is computed on the upper soil layer (30 cm). Weather data (daily temperature and precipitation) used as input are the values of the analysis grid ERG5 on the Emilia-Romagna region.

Results and Discussion
The main output of RainBO has been the integration of the data and models described in the Materials and Methods encapsulated in an organic platform [47], as presented in the next paragraph.

The RainBO Platform
The RainBO platform consists of the following key elements: • database containing monitoring, territorial, and historical data • software modules, which are the platform intelligence • graphic interface, which is the platform output One of the most important features of the platform is the database containing the monitoring data, whose functionality is to integrate data collected from different monitoring infrastructures, both conventional and unconventional, as well as forecast data, hydrological, and meteorological models, and estimated ones.
In particular, the implementation of an advanced monitoring infrastructure within the RainBO Life project consists of the integration of these types of data: • real sensors data (e.g., weather stations) • "virtual sensors" data, not associated with observed measurements from physical sensors, but obtained indirectly through the estimation of correlated data or from simulation models • forecast data, provided by simulation models This structure allows us to monitor extreme precipitation events, their evolution, and to generate early warnings. It is worthy to mention that the concept of "virtual sensor" allows us to integrate information from observed and not observed data sources, georeferencing them with the same reference system and synchronizing them over time.
The integration of these new virtual sensors into the RainBO platform has been accomplished in a simple way, using the same data model defined for the physical sensors, without any extension or specialization and providing the platform with an enhanced monitoring infrastructure.
The territorial database hosts both the input data and the output data coming from the processing of the application modules, as well as, obviously, the data necessary to describe the territorial characteristics.
The RainBO platform architecture has been designed according to the following attributes: • open: each module exposes standard interfaces (web services) to ensure system generality and replicability as well as interoperability and integration with other platforms • centralized: each DB is centralized and enables data sharing, managed, and updated by different users • scalable: each module is developed so as to be implemented on different machines • modular: the platform is formed by individual modules ensuring more flexibility, maintainability over time, as well as platform evolution as each module can evolve or be replaced independently from each other • configurable: each module is configurable, i.e., the operating parameters must be read from the table and not written in code RainBO platform has been conceived according to two operational modes: 1. Planning support 2.
Event management Both the modes display a time bar with different time ranges according to the selected operational mode: the menu bar for the planning support mode refers to historical events whereas the menu bar for the event management mode refers to monitoring data (from −24 h to +72 h), in addition to menu bars and GIS maps specifically for each mode ( Figure 11).

657
To collect territorial data, the data model was defined according to the existing national and 658 regional guidelines for civil protection emergency plans. It requires the mapping of all the critical, 659 sensible, and strategic items, flood maps, river and territory maps (e.g., land use, network 660 infrastructures, buildings, factories, parks).

661
The territorial database of the RainBO platform contains the existing data listed above: basic 662 regional maps, hazard maps from the Floods Directive, and territorial data at municipality level (e.g.,

663
hospitals, schools, emergency areas). The territorial database also contains the maps resulting from 664 the vulnerability module elaborations.

665
The definition of the spatial data model required an important standardization work to specify 666 its structure, name, and format, to define a standard at the implementation level, both for the RainBO 667 platform and as a reference for data coming from third-party systems.

668
For instance, the attributes to be provided for schools are the location, the polygon of the area,

Territorial Data
To collect territorial data, the data model was defined according to the existing national and regional guidelines for civil protection emergency plans. It requires the mapping of all the critical, sensible, and strategic items, flood maps, river and territory maps (e.g., land use, network infrastructures, buildings, factories, parks).
The territorial database of the RainBO platform contains the existing data listed above: basic regional maps, hazard maps from the Floods Directive, and territorial data at municipality level (e.g., hospitals, schools, emergency areas). The territorial database also contains the maps resulting from the vulnerability module elaborations.
The definition of the spatial data model required an important standardization work to specify its structure, name, and format, to define a standard at the implementation level, both for the RainBO platform and as a reference for data coming from third-party systems.
For instance, the attributes to be provided for schools are the location, the polygon of the area, the number of students, employees and disabled persons, the number of floors of the building, the contacts of the school manager, including the phone numbers. The set of this information, as will be explained below, is necessary both to define the degree of vulnerability of critical sites, but also for the set up of the early warning system.

Hazard, Vulnerability, and Risk Maps
The RainBO platform includes reference hazard maps as: • hazard maps deriving from the Floods Directive: they represent the potential extent of flooding caused by (natural and artificial) watercourses or by sea, with reference to three scenarios (rare floods (P1-L), infrequent (P2-M) and frequent (P3-H)) represented with three different shades of blue, where the decrease of frequency of flooding corresponds to the decrease in intensity of color. The Floods Directive hazard maps derive from the national hydro-geological management plan (PAI) and they are available for the main basins • hazards maps from specific hydraulic model/studies for small basins, not included in the Floods Directive maps • historical events maps that are maps of the flooded areas due to past events. These maps represent an important source of additional information to compare reference maps listed before and real ground effects expected in case of an event.
Concerning these maps, Figure 12a shows the hazard map related to the Parma river in the city of Parma from the Floods Directive. Figure 12b shows the same map integrated with the flooded area during the flood of October 13, 2014. This additional information extends the standard risk area to a secondary one.

687
Concerning these maps, Figure 12a shows the hazard map related to the Parma river in the city 688 of Parma from the Floods Directive. Figure 12b shows    The vulnerability module calculates the degree of vulnerability as described in Section 2.3.3. As a result, 32 vulnerability maps, corresponding to 32 predefined time frames, are produced.
Furthermore, vulnerability maps are a support for territorial planning in prevention and preparedness stages also according to the regional law of December 21, 2017, n.24, which requires the preliminary assessment of the risk, and therefore of the vulnerability, with respect to different types of events, including the hydraulic one, for the purpose of defining the regional urban plan.
In case of an ongoing event, or forecast event, the RainBO platform can select and make available the vulnerability map of the corresponding time frame, providing support for its management.
In more detail, the main difference between the 32 vulnerability maps concerns the distribution of residents, workers, and users of sensible targets during the whole day. By way of example, during the night sensible targets, workplaces, and facilities are usually closed, therefore it is supposed that most of the population are in residential areas. As a consequence, in the vulnerability map referring to this time frame, the urban residential areas are represented in red, whereas during the morning of a working day these areas are green; during the night frame, also sensible targets as schools, gyms, or museums are green areas, whereas during opening hours these areas are red. Figure 14 shows the vulnerability map of an area of the city of Bologna corresponding to a working day at 4 a.m. The vulnerability level shown in the following pictures is: • Red for high vulnerability, which means the presence of many people in the grid cell.

•
Orange is medium-high vulnerability • Yellow is medium vulnerability • Green is low vulnerability, due to the presence of few people in the grid cell As a result, the RainBO platform provides also the corresponding 32 risk maps. In case other up-to-date or detailed hazard maps (besides the Flood Directive maps) are available, the platform allows us to specify the hazard map on which the risk can be calculated.
As described above, RainBO risk maps based on the vulnerability maps are more detailed and focused on the distribution of sensible and strategic targets and on the time depending presence of citizens in an urban area, with respect to Floods Directive data.
As a result, in case of a forecast event, the risk map, derived from the combination of the hazard map (now from the Floods Directive) and the vulnerability map (selected according to the time frame by the timing of forecast alert between the 32 maps of default) could better support decision-makers to prioritize the warning from the red areas to the green ones.
Moreover, in the planning phase, the capacity of the software to calculate new risk maps from vulnerability maps can support municipal technical and planning offices to evaluate territorial planning choices, as a preventive measure.
most of the population are in residential areas. As a consequence, in the vulnerability map referring Figure 14 shows the vulnerability map of an area of the city of Bologna corresponding to a 721 working day at 4 am. The vulnerability level shown in the following pictures is:

722
• Red for high vulnerability, which means the presence of many people in the grid cell.
to-date or detailed hazard maps (besides the Flood Directive maps) are available, the platform allows 733 us to specify the hazard map on which the risk can be calculated.  . Vulnerability map at 04:00 of a working day. In the color legend, green ("vulnerabilità 1") is low vulnerability, yellow ("vulnerabilità 2") is medium vulnerability, orange ("vulnerabilità 3") is medium-high vulnerability, red ("vulnerabilità 3") is high vulnerability. Figure 15 shows the risk map at 4:00 a.m. of a working day related to the Ravone creek, produced by using the hazard map of the catchment integrated within the RainBO platform and not included in the Floods Directive.   759 760 Figure 15. Risk map at 04:00 of a working day referred to the Ravone creek. In the color legend, green ("rischio 1") is low risk, yellow ("rischio2") is a medium risk, orange ("rischio3") is medium-high vulnerability, red ("rischio3") is high risk.

Historical Events
RainBO platform integrates a database where the most significant past flood events on test catchments provided by Arpae are stored. This database is aimed at collecting meaningful information for the management of future events. The time bar allows us to identify and explore the flood events in the planning support mode. The most important information linked to the event (e.g., data catchment of interest, the exceeded threshold of water level gauge, the description of flooded area uploaded on the platform as vectorial data) are highlighted in one view Figures 16 and 17. catchments provided by Arpae are stored. This database is aimed at collecting meaningful

Observed, Forecast, and Estimated Data
With regard to the observed data, the RainBO platform collects data from heterogeneous sensors: inclinometers for landslide monitoring, rain gauges, and water level gauges for hydro-pluviometric monitoring of the Arpae network (including the new monitoring stations installed for the RainBO project), inductive-loop detectors for traffic monitoring. The RainBO platform integrates more than 1500 sensors of different types and technologies ( Figure 18).   With regard to the estimated data, the system displays maps of radar reflectivity ( Figure 20) and radar-estimated precipitation. These two variables are directly related. The hourly maps of these variables can be explored using the time bar for the previous 24 h and it allows the monitoring of ongoing events.
Concerning CMLs, the RainBO platform integrates 73 microwave virtual sensors, corresponding to the midpoint of radio links on the Lepida wireless network (Figure 21).

812
Therefore the hydraulic forecast shows the same time shift. (Figure 22b).

Hydrologic Forecast
RainBO platform provides the forecast of exceeding critical thresholds on the two test catchments, the River Parma for the Parma municipality and the creek Ravone for Bologna municipality. The forecast refers to specific hydraulic sections of the catchments: the culvert entry for Ravone creek, Ponte Verdi, and Ponte Nuovo hydraulic sections for the Parma River.
The forecast covers different time ranges, 6 h for the Parma and 72 h for Ravone, as it is produced by different forecast models, as explained before. Concerning the Parma case study, in 2017, December 12th, the Parma river basin was interested in an intense and prolonged rainfall event, starting from the 14:00 UTC of the 10th of December a weak rainfall interested the basin till the day after when the rainfall started growing in intensity and space. The peak in terms of mean area rainfall was at midnight on December 11th with about 8 mm/hr ( Figure 23).

828
Looking at the riverside we observed, at 05:00 UTC, the first raise of levels in the river Parma at

829
Ponte Verdi with a level 1 threshold crossing event on the same day at 11:00 UTC. Levels raised 830 continuously in the next hours due to the prolonged rainfall in the basin, reaching the threshold 2 at Looking at the riverside we observed, at 05:00 UTC, the first raise of levels in the river Parma at Ponte Verdi with a level 1 threshold crossing event on the same day at 11:00 UTC. Levels raised continuously in the next hours due to the prolonged rainfall in the basin, reaching the threshold 2 at 16:00 UTC and a peak of 3.14 m (meter above local reference) at 03:00 UTC of the 11th of December, and at the same time, the rainfall in the basin ended with a total amount of mean area rainfall for the entire event of about 116 mm.
The forecast probability generated by the RF model during this event is shown in Figure 24. We can notice that all threshold reached a high probability, between 90-100%, this is obviously related to the high level reached by the river during the event, but we can also consider the forecast lead time provided by the model, in fact, if we look at Table 2, we can find a good consistency between probability and observed effects few hours later, especially for threshold 3 where at 18:00 UTC the probability raise from 12% to 44%, and lately at 20:00 reached a 73%, about 6 h before the effective threshold crossing event, observed at 02:00 UTC (red color).

852
In more detail, in the event management mode, two specific icons conceived for early warning 853 are included. The icons are highlighted if a warning occurs. The first icon refers to ongoing events 854 (alert is triggered by observed data of threshold exceeding recorded by water level gauges), whereas 855 the second icon refers to forecasts (alert is triggered by hydrologic models).

856
In the presence of an alert, a specific dashboard is opened by clicking on the reported event. The

Early Warning Module
An early warning module was developed in the RainBO platform in order to identify the expected scenario as a function of monitoring, forecast and vulnerability maps. A filtering and signaling system allows to select and group essential information useful for the users during extreme events.
In more detail, in the event management mode, two specific icons conceived for early warning are included. The icons are highlighted if a warning occurs. The first icon refers to ongoing events (alert is triggered by observed data of threshold exceeding recorded by water level gauges), whereas the second icon refers to forecasts (alert is triggered by hydrologic models).
In the presence of an alert, a specific dashboard is opened by clicking on the reported event. The dashboard includes only the essential and useful information for the Event management, as the list of catchments where critical thresholds are exceeded (Figure 25), the link to the corresponding charts (Figure 26), the vulnerability map corresponding to the current or forecast scenario. Furthermore, in the event management mode, a specific tool allows selecting an area of interest where critical targets and sensible targets (with users) are listed, so that reference persons can be identified in order to contact them ( Figure 27).

890
Another EU Interreg project named RAINGAIN sought to obtain detailed rainfall data at an 891 urban scale, to predict urban flooding and to implement the use of rainfall and flood data in urban 892 water management practice. In this case, radar technology is used to provide estimates of rainfall in 893 time and space in cities.

894
RainBO provides estimates of precipitation from different sources such as rain gauges, radar 895 data and CML, and provides a forecast of precipitation from a limited area meteorological model.

896
These data related to precipitation are integrated with other processes (e.g., hydrological modeling) 897 and with territorial data useful before and during the event.

898
Furthermore, the EU FP7 project STAR-FLOOD aimed to improve the implementation of flood 899 risk strategies in urban areas by building appropriate and resilient flood risk governance. This project 900 stressed the need to have proactive and risk-based systems for forecasting, warning and emergency 901 responses for urban floods that also need to use collaborative approaches. Therefore STAR-FLOOD 902 is more focused on the planning phase, whereas RainBO comprises also a consistent part devoted to 903 forecasting and event management.

904
Finally, the EU Horizon 2020 project FLOOD-serv aimed to develop a collaborative platform 905 among citizens, public authorities and other stakeholders, which enable alerts in due time to reduce 906 the adverse effects of the floods. In this regard, it is worthy to mention that the RainBO platform is

Discussion
In general terms, the flood risk in urban areas can be reduced through the implementation of planning, and monitoring and forecasting phases. These activities are related to different time scales and can be supported through a tool that uses multiple information collected and which can display them easily.
The RainBO platform integrates meteorological and territorial data in time/space and provides a holistic interface for the planning phase and disaster risk reduction, allowing to prevent impacts of flood events and reduce the residual risk during their occurrence.
As flooding in urban areas in Europe is increasing, European projects tackled this issue in urban environments with different features. The EU Interreg project URBAN-PREX aimed at monitoring/forecasting and developing online public EWS for extreme precipitations and pluvial floods in urban areas. URBAN-PREX focused on precipitation monitoring by means of rain gauges located in urban areas and precipitation forecasts.
With respect to this project, the approach of RainBO includes further modules that deal with the planning, monitoring, and forecasting phases. Maps of hazard, vulnerability, and risk are available in the planning phase; data and innovative technology such as CML have been included to monitor precipitation events in the monitoring phase; finally, there is a very clear focus on specific hydrological modeling in the forecasting phase.
Another EU Interreg project named RAINGAIN sought to obtain detailed rainfall data at an urban scale, to predict urban flooding and to implement the use of rainfall and flood data in urban water management practice. In this case, radar technology is used to provide estimates of rainfall in time and space in cities.
RainBO provides estimates of precipitation from different sources such as rain gauges, radar data and CML, and provides a forecast of precipitation from a limited area meteorological model. These data related to precipitation are integrated with other processes (e.g., hydrological modeling) and with territorial data useful before and during the event.
Furthermore, the EU FP7 project STAR-FLOOD aimed to improve the implementation of flood risk strategies in urban areas by building appropriate and resilient flood risk governance. This project stressed the need to have proactive and risk-based systems for forecasting, warning and emergency responses for urban floods that also need to use collaborative approaches. Therefore STAR-FLOOD is more focused on the planning phase, whereas RainBO comprises also a consistent part devoted to forecasting and event management.
Finally, the EU Horizon 2020 project FLOOD-serv aimed to develop a collaborative platform among citizens, public authorities and other stakeholders, which enable alerts in due time to reduce the adverse effects of the floods. In this regard, it is worthy to mention that the RainBO platform is addressed only to decision-makers; decision-makers of a municipality can support both territorial planning and early warning systems through this platform.
RainBO platform has the added value of combining in only one integrated tool territorial data, historical data, real-time monitoring, a crowdsourcing system, hydraulic models for small creeks and medium rivers and the early warning system.
The strengths of the project both in terms of the development of new products and enhancements of existing ones are: • Vulnerability map calculation module (for hydraulic risk purposes) • Integration of observed, estimated, and predicted data • Hydrological simulation model for small basins Moreover, a first operational level has been defined for a rainfall monitoring system based on CML and it consists of satisfactory reliability in monitoring convective events, mainly during the summer and spring seasons.

Conclusions
The RainBO platform is open, centralized, scalable, modular and configurable; it means that it can include various data and models in a different time and space scale, providing an opportunity of replicability and scalability in other geographical and administrative contexts and also for different purposes.
Furthermore, the effectiveness of the RainBO platform has been confirmed by the other municipalities in the Emilia-Romagna Region (i.e., Cento, Comuni della Val Samoggia, Anzola dell'Emilia) keen to demonstrate this platform.
Some limitations of the RainBO platform are to be mentioned: • The platform is addressed to decision-makers but not to citizens (the citizens can only send information through crowdsourcing) • The hydrologic models used within the platform have to be calibrated and validated if applied in different river basins, therefore they are not easily replicable in other contexts • A very large amount of information can be challenging to manage From a future perspective, the extension of this platform can be foreseen in other urban areas due to the diffusion of open data. As an increasing number of territorial open data is now available, one of the benefits of the platform is also the convergence on the same hub of these data organized in a functional and rational way.
The specific features of this platform allow the upload of datasets provided from international programs (e.g., Copernicus) to supply the lack of local information or infrastructures.
Another future development could be the re-engineering of the platform with a simplified set of information, conceived for the citizens of specific urban areas.
In the future, an effort should be made to enhance the robustness of this platform, i.e., improving the operational functioning of the monitoring and forecasting phases.