1. Introduction
In the last two decades, significant efforts have been made to predict flash floods and their consequences. These events are characterized by intense rainfall occurring in small geographical areas or subcatchments (Rebora et al. 2013; Braud et al. 2016; Edouard et al. 2018; Faccini et al. 2021; Guastavino et al. 2022). Despite notable advancements in hydrometeorological forecast systems in recent years (Buzzi et al. 2014; Davolio et al. 2015; Lagasio et al. 2019; Ramos Filho et al. 2021; Wang et al. 2021; Lagasio et al. 2022b), flash floods are still widely recognized for their inherent challenges in terms of predicting their location, intensity, and spatiotemporal structure of precipitation.
The idea of using meteorological data, obtained from both ground-based and remote sources, together with various three-dimensional schemes of rainfall prediction systems, has emerged within the hydrometeorological community. There is a consistent focus on enhancing and testing nowcasting radar extrapolation techniques (Panziera et al. 2011; Foresti et al. 2016; Imhoff et al. 2020, 2022), machine learning techniques from ground stations data (Pirone et al. 2023), and assimilation methods in numerical weather prediction models (Lagasio et al. 2019; Gastaldo et al. 2022) for improving results. Recently, Zhang et al. (2023) introduced a method combining physical evolution schemes and conditional learning methods to improve nowcasting skill. However, improving extreme precipitation nowcasts is still a challenge. In general, the prevailing approach is to update forecasts typically focusing on short time ranges frequently. Bowler et al. (2006) developed a short-term rainfall forecast system by combining radar extrapolation techniques and the NWP, assigning appropriate weights based on the extrapolation and NWP forecasts. Hwang et al. (2015) defined a merged rainfall field by using a method that preserves the features from multiple images considering saliency (or importance) as the locations of strong cells. Atencia et al. (2020) presented a technique that allows for keeping the spatial correlation of the errors to merge the forecast from Lagrangian extrapolation and mesoscale models with locally different weights to enhance rainfall predictions, while Franch et al. (2020) explored a deep learning approach to improve extreme rainfall events. The work of Sideris et al. (2020) developed a blending method that combines extrapolation techniques and the NWP using a linear weighting function calibrated based on the performance of NWP. This weighting function varies in time instead of being constant as in simpler approaches.
Many efforts have been made to harness nowcasting and short-term rainfall predictions to enhance hydrological forecasting, yielding promising results. Berenguer et al. (2005) validated a radar-based nowcasting technique in terms of hydrological simulations. Silvestro and Rebora (2012) pursued a similar approach but employed a probabilistic method. Liguori et al. (2012) investigated the feasibility of combining an extrapolation radar rainfall forecast with a high-resolution NWP rainfall forecast to evaluate the flow predictions in urban areas. More recently, Poletti et al. (2019) and Charpentier-Noyer et al. (2022) have conducted an analysis on the hydrological response anticipation driven by various rainfall fields products, which include blending methodologies and data assimilation frameworks. These studies have also demonstrated the benefits of utilizing both extrapolation and NWP in a blended rainfall product compared to using extrapolation alone. Heuvelink et al. (2020) employed precipitation nowcasting to predict streamflow in the Netherlands and found that uncertainty increases when catchments dimension decreases. Lovat et al. (2022) presented and discussed the hydrometeorological verification of systems in relation to heavy rainfall events in the Mediterranean environment.
In this work, a further investigation into the aforementioned topics is carried out with the aim of enhancing short-term forecasts and frequently updating streamflow predictions by utilizing recent and state-of-the-art systems. Two methodologies to perform rainfall short-term forecasts are considered. The first methodology is an improved version of the blending method described in Poletti et al. (2019). A more robust assimilation technique is used in the NWP system (Lagasio et al. 2019), and the blending function used for linear combination is made more flexible and time dependent (Sideris et al. 2020). The second methodology, described in Lagasio et al. (2022a), also utilizes the same components (NWP and nowcasting method). However, it takes advantage of the overlapping forecast time windows from different NWP runs when applying a frequent assimilation approach.
Eventually, a distributed hydrological model is employed to forecast streamflow and serves as rainfall integrator. This approach addresses the “double penalty” problem (https://www.ecmwf.int/en/about/media-centre/science-blog/2023/verifying-high-resolution-forecasts, last accessed 29 January 2024) encountered in many meteorological verification approaches. By utilizing the distributed hydrological model, the results are evaluated in terms of catchment-scale processes.
A hindcast experiment is then conducted, encompassing a significant number of basins (59) and two periods of approximately 10 days each, characterized by a high frequency of events that predominantly affected northwest region of Italy. Due to a combination of factors such as (i) the frequent update of forecasts, (ii) the duration of the hindcast periods, and (iii) the inclusion of multiple basins, a substantial dataset was available for assessing how the set of evaluation scores used to measure performances varies with the forecast lead time. The initial analysis focuses on comparing the potential assimilation of both radar and lightning data versus radar data alone using only the first nowcasting method described in Poletti et al. (2019). Subsequently, a comparison is made between the two nowcasting methods, namely, the one proposed by Poletti et al. (2019) and the method presented in Lagasio et al. (2022a).
One interesting aspect is the fact that the two short-term rainfall forecast systems exploit the comparison of forecast and observed rainfall in order to dynamically adjust their setting and parameterization. The objective of this research is to determine whether one strategy is better than the other when predicting streamflow. Another novel aspect is the investigation of the effects resulting from the utilization of different datasets in the assimilation process. Both of these topics are approached in terms of streamflow prediction.
2. Materials and methods
a. Study area and data
The study area is located in the northwestern part of Italy and includes watersheds belonging mainly to the Liguria region, as well as areas of southern Piedmont and northwestern Tuscany (Fig. 1). This region is characterized by a Mediterranean climate, with the main rainy season occurring during autumn. The area is prone to highly intense rainfall events that often result in floods and flash floods, occurring with a certain frequency (Silvestro et al. 2016; Parodi et al. 2020). The morphology of the study area is characterized by basins ranging in size from approximately 10 to 104 km2. These basins exhibit steep slopes, and a significant portion of the territory is covered by forests and brush. The region features mountains having altitudes reaching up to 2000–2500 m. Cities and towns are predominantly situated along the coast or adjacent to the rivers (Brandolini and Spotorno 1997; Brandolini and Sbardella 2001; Roccati et al. 2018; Faccini et al. 2021). As a result, the catchments’ response time typically ranges from 1 to 12 h (Delrieu et al. 2005; Silvestro et al. 2018). During the early autumn, the temperatures along the coast commonly range between 20° and 25°C, which gradually drop to 8°–12°C in the late autumn. Generally, the snow accumulations are negligible in this season, except for some small areas situated at altitudes higher than 2000 m.
(a) Hydrological streamflow simulation system operationally implemented at Italian National Scale available on the Dewetra Early Warning Support system. Markers with violet borders are the hydrological sections where streamflow observations are available. Squares represent sections where streamflow has no trend, and triangles represent the ones where trend is increasing or decreasing (trend is estimated on a time window of 12 h centered on the time of most recent run of the model). The zoom shows the area of interest for the analysis. (b) Higher detail of the 59 sections involved in the analysis (red markers), the hydrographic network (blue lines), and the main basins (green polygons). The digital elevation model shown in grayscale highlights the morphology.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
Radar data used in this study are obtained from the Italian National Radar Network (Vulpiani et al. 2012, 2016) and provided by National Civil Protection (https://dpc-radar.readthedocs.io/it/latest/, last accessed 26 January 2024). The following products are used:
-
surface radar intensity (SRI): this product serves as input to the nowcasting system (frequency of use 10 min). The National Civil Protection delivers an adjusted version of SRI (SRIadj), which incorporates rain gauges observations. The SRIadj is the product used in the experiments presented.
-
constant-altitude plan position indicator (CAPPI) of radar reflectivity (CAPPIZ): CAPPIZ at three different heights (2000, 3000, 5000 m) is used in the data assimilation process every 3 h for the three-dimensional variational assimilation system (3DVAR) application.
Data obtained from the Italian National Meteorological Network are utilized to generate the required inputs for streamflow simulations. Hourly temperature, air relative humidity, solar radiation, and wind velocity data are employed after undergoing spatial interpolation (DPC 2022). Hourly rainfall data, along with radar data, are combined using a merging algorithm derived from Sinclair and Pegram (2005) and described in Bruno et al. (2021).
Lightning data, along with radar CAPPI, are utilized for assimilation purposes. They are assimilated using a 3DVAR and a nudging technique at the frequency outlined in section 2d. For this purpose, SFLOC (term derived from the words spheric and location) data of the lightning flashes are obtained from the LAMPINET Aeronautic Meteorological Service, which is based on Vaisala technology (De Leonibus et al. 2010). The LAMPINET network consists of 15 sensors with improved accuracy from combined technology, offering enhanced sensitivity and performance, uniformly distributed across the Italian national territory (De Leonibus et al. 2010). The network employs both MDF (magnetic direction finding) and TOA (time of arrival) techniques (De Leonibus et al. 2008). Network demonstrates a detection efficiency of 90% for peak currents exceeding 50 kA with a location accuracy of about 500 m throughout Italy. The sensors used are of the IMPACT-ESP (Improved Accuracy through Combined Technology Enhanced Sensitivity and Performance Sensor) type. The sensors detect radio frequency energy from both cloud-to-ground (CG) and intracloud (IC) discharges within a bandwidth ranging from 1 to 350 kHz. In addition to location accuracy, the sensors also measure azimuth angle, time of signal arrival, peak signal strength, and discharge width.
b. Streamflow short-term forecast framework
The short-term forecast framework is made by a cascade of models that use observations as input or for their initialization:
-
an NWP with frequent cycles of the 3DVAR assimilation process, described in section 2d;
-
a nowcasting algorithm based on rainfall radar fields, described in section 2e;
-
two algorithms for combining rainfall forecasts derived by point 1 and point 2, described in sections 2f and 2g; and
-
a hydrological model used to transform rainfall forecasts into streamflow forecasts, described in section 2c.
For a general scheme one can refer to Fig. 5 of Poletti et al. (2019).
c. Hydrological model
The continuous and distributed hydrological model employed in this experiment is extensively described in Silvestro et al. (2013, 2015, 2021) and is referred to as Continuum. The specific configuration adopted in this study is widely documented in Davolio et al. (2017) and Silvestro et al. (2018), which also describe the calibration strategy. Furthermore, Bruno et al. (2021) presents the implementation of the hydrometeorological system at the Italian National Scale for civil protection purposes. The model utilized in this work has a spatial resolution of 0.05° (around 480 m) with a time resolution of 1 h. The system is implemented nationwide and incorporates over 350 control sections (Fig. 1). Streamflow time series from these control sections are utilized to monitor and forecast the intense rainfall events impacts. The results are published on myDewetra, one of the official Early Warning Support Systems employed by Italian Civil Protection (Pagliara et al. 2011; https://www.mydewetra.org/wiki/index.php/FloodPROOFs_Italia_Deterministico_-_Osservazioni).
As described in section 3, the analysis periods under consideration primarily focus on the northwest region of Italy. Consequently, 59 sections within this area are selected to conduct the experiments and the analysis. The drainage areas of these selected sections range from 10 to 103 km2. To initialize the state variables of the Continuum model at the beginning of the two analysis periods, a long-term run is performed using observations as input.
d. WRF and assimilation approach
The Weather Research and Forecasting (WRF) Model v3.8.1 (Skamarock et al. 2008) was chosen as the numerical weather model for this study. The WRF Model is a fully compressible nonhydrostatic model that utilizes mass-based terrain-following coordinates. It was developed by the National Center for Atmospheric Research (NCAR) in collaboration with several institutes and universities. The model serves both for operational weather forecasting and research purposes. For this work, three two-way nested domains (Fig. 2) were selected. The horizontal grid spacing for each domain (Fig. 2) is 22.5 km (216 × 191 grid points), 7.5 km (523 × 448), and 2.5 km (430 × 469). All domains consist of 50 vertical levels and extend up to 50 hPa.
WRF Model configuration. The three boxes represent the horizontal grid spacing domains: 22.5 km (external domain), 7.5 km, and 2.5 km (smaller domain that includes mainly Italy).
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
All the simulations in this study employ the same set of physical parameterizations which have been previously validated and proven successful in other studies focusing on extreme events in Italy (Fiori et al. 2017; Lagasio et al. 2017, 2019, 2022b; Mazzarella et al. 2022). The following parameterizations are adopted:
-
Surface layer: the MM5 scheme is adopted (Paulson 1970; Dyer and Hicks 1970; Webb 1970).
-
Surface fluxes: a parameterization scheme following Beljaars (1995) is used to enhance surface fluxes of heat and moisture.
-
Soil model: the Rapid Update Cycle (RUC) scheme is used as a multilevel soil model (6 levels), with higher-resolution levels distributed in the upper soil parts (0, 5, 20, 40, 160, and 300 cm). This soil model solves the heat diffusion and Richards moisture transfer equations (with a layer approach) and, in the cold season, considers phase changes in soil water (Smirnova et al. 1997, 2000).
-
Planetary boundary layer (PBL): The dynamics are parameterized with the diagnostic nonlocal Yonsei University (YSU) PBL scheme (Hong et al. 2006).
-
Microphysics: The WSM6 six-class scheme is adopted (Hong and Lim 2006).
-
Radiative processes: They are parameterized by means of the longwave and shortwave RRTMG schemes (Iacono et al. 2008).
-
Convective parameterization: the new simplified Arakawa–Schubert scheme (Han and Pan 2011) is adopted, except in the inner domain where convection is explicitly resolved.
1) Assimilating radar reflectivity data
The forecasts with data assimilation (WRF-DA setup) were conducted using a 3-h cycle 3DVAR technique of radar reflectivity observations as indicated in Fig. 3. The assimilation is denoted as “DA hh REFL,” where hh represents the assimilation time instant. The 3DVAR assimilation purpose is to minimize the difference between background forecast and the observations by optimizing a cost function (Ide et al. 1997). The WRF data assimilation package WRFDA (Barker et al. 2012) v3.9.1 is employed. In this study, the reflectivity modified direct operator as presented in Lagasio et al. (2019), is used. The control variable option selected is 5 (CV5), which refers to the
Assimilation timing scheme for radar and lightning data.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
The NMC method was applied over the entire month of October 2015, with a 24-h lead time for the forecasts starting at 0000 UTC and a 12-h lead time for the ones initialized at 1200 UTC on the same day. The differences between the two forecasts (t + 24 and t + 12) valid for the same reference time were utilized to calculate the domain-specific error statistics.
2) Assimilating lightning data
The mixing ratio qυ described in Eq. (1) is computed at each LDA time step for grid points where flashes are observed, indicated by X greater than zero, within the layer between the 0° and −25°C isotherms. The computed mixing ratio from Eq. (1) is then compared with the predicted mixing ratio by the model. If the computed mixing ratio of Eq. (1) is larger than the simulated one, the modeled mixing ratio is replaced with the value obtained from Eq. (1). On the other hand, if the modeled mixing ratio is already larger, it remains unchanged. It is important to note that this method can only introduce additional water vapor to the forecast.
In this study, the lightning assimilation is conducted according to the timing presented in Fig. 3, denoted as “DA LIGHT.hh1-hh2,” represents the assimilation time coverage, starting from hh1 and extending until hh2.
e. PhaSt nowcasting algorithm
To overcome the method limitations, which maintain constant precipitation volumes based on the most recent observation, the rainfall volume correction method described in Poletti et al. (2019) is implemented in this study. This correction method involves determining the percentage increasing or decreasing of the precipitation total volume for each time step relative to the time of the most recent radar observation available, using information derived from NWP. This percentage adjustment is then applied to the PhaSt nowcasted fields (refer to Fig. 3 in Poletti et al. 2019). This approach allows us to utilize information derived by the NWP but while still leveraging the predictive capabilities of the extrapolation techniques, particularly during the first 1–3 h of forecast.
f. Blending approach
The blending strategy proposed in this study falls under the category of linear combinations of NWP and nowcasting rainfall fields (Hwang et al. 2015). Building upon the methodology presented in Poletti et al. (2019), the technique enhanced by incorporating the approach utilized by Sideris et al. (2020). This approach involves utilizing the comparison between models and observations to parameterize the blending function. As a result, the nowcasting and the most recent NWP run are dynamically weighted based on their performance, allowing for adaptive blending.
A similar approach is applied to estimate ρPhast, but since PhaSt can be run with higher frequency (FrPhaSt = 10 min), the correlation is estimated for all the predictions made in the time interval from tr = tnow − 2.5 h to tr = tnow − 2 h. The correlation ρPhast is then estimated as average of those values.
In this analysis, the blended final rainfall field is generated with an update frequency of 1 h. Figure 4 provides a schematic representation of the blending framework. In the presented study, NWP is updated every 3 h, but nowcasting can be updated with high frequency as well as the a parameter estimation.
Blending framework scheme and forecasts set used to estimate the correlations between WRF, PhaSt, and observations. PhaSt is updated every 10 min, WRF with data assimilation is updated every 3 h, and blending is updated every hour.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
g. SWING method
The postprocessing algorithm SWING (score-weighted improved nowcasting; Lagasio et al. 2022a) is developed to account for the temporal and spatial uncertainty in simulating convective fields. SWING is employed considering the three forecasts that span the same 6-h time window, repeated every 3 h (refer to Fig. 3 in Lagasio et al. 2022a). It combines the PhaSt output and the WRF Model simulations to generate an updated forecast every 3 h. However, in this particular case, SWING is not used to produce the rainfall hazard scenarios presented in Lagasio et al. (2022b). Instead, it provides hourly weighted rainfall every 3 h obtained from the three simulations that cover the same 6-h period of forecast (Fig. 3 in Lagasio et al. 2022a) to provide a useful input to the hydrological model. The forecast weighing is determined by evaluating the simulation performance during the 3 h preceding the time of interest using an object-based approach. This evaluation involves comparing the modeled rainfall objects with the observed one obtained by the radar and rain gauges rainfall data on the national map (Bruno et al. 2021). The rainfall objects are identified by isolating the regions with rainfall values exceeding 2 mm. The comparison between the observed and modeled rainfall objects involves evaluating various objects’ characteristics, such as rainfall volumes, areas extension, distance between center of gravity, objects’ orientation, and objects’ intersection. These characteristics are used to calculate scores and determine the correspondence between the observed and modeled objects. The scores are then summarized into an overall field score (OFS; Lagasio et al. 2022a). In addition to the OFS, a missed object score (MOS) is calculated to account for observed objects that are not present in the model. The MOS ranges from 0 to 1 (Lagasio et al. 2022a). These scores, along with the OFS, are combined to obtain a final reliability score (RS; Lagasio et al. 2022a). The RS score is used to weight the three considered forecasts, resulting in the mean rainfall fields that can be utilized as input for the Continuum hydrological model.
3. Hindcast experiment and hydrological verification
a. The experiment
The hindcast experiment involved simulating the entire forecast system for two distinct periods. The first period spanned from 14 to 22 October 2019, while the second period encompassed 14–21 November 2019 [refer to Lagasio et al. (2022a) for detailed information on the two periods]. For sake of simplicity, they will be named October and November. According to official event analysis documents (https://www.arpal.liguria.it/tematiche/meteo/pubblicazioni-bis/rapporti-di-eventi-significativi/rapportieventisignificativi-2019.html), both periods were characterized by nearly continuous rainy weather and various flood events. During the October period, the events were mainly concentrated in small areas, with local rainfall accumulations of up to 100–120 mm in 1 h and 250–350 mm in 6 h. On the other hand, the events in November were more spatially distributed (Molini et al. 2009) with significant rainfall accumulations also on larger areas (50–100 mm in 24 h and 103 km2), as well as local storms with hourly accumulation of 40–60 mm and local accumulations of 300–400 mm in 24–36 h.
Each period consists of approximately 7–8 days, and the two presented methods have different update frequencies:
-
The blending-driven method, which utilizes the most recent WRF run, is updated every hour.
-
The SWING-driven method is updated every 3 h.
The maximum lead time of rainfall forecast is 6 h. However, for hydrological analysis, from hour 7 to 12 zero rainfall is considered. This is because the larger basins involved in the experiment have drainage area on the order of 103 km2, with concentration times around 10 h.
Therefore, a large number of forecasts is available to conduct a robust analysis. However, for the blending approach, there is a larger number of samples compared to the SWING approach. As a result, an initial investigation is conducted to compare the simulation using both radar and lightning assimilation or radar-only assimilation using all the available blending forecasts. The comparison with the SWING method is performed only when both methods are available. Each method is applied using the two assimilation configurations: the first one uses radar data only, while the second uses both radar and lightning data.
The benchmark used for the comparison is the hydrological model run using observed meteorological variables as input. This approach, which has been employed in previous studies (Borga 2002; Vieux and Bedient 2004), allows for the inclusion of multiple basins in the analysis while also eliminating the potential effects of hydrological model calibration errors. To evaluate the performance scores for different forecast lead times, the method described in Berenguer et al. (2005) is utilized.
b. Performance indicators
Table 1 provides an overview of the three basin classes, along with a rough estimation of the range of concentration times for each class.
Basin classes defined basing on drainage area together with an approximated indication of response time.
A further analysis was carried out to assess the capability of predicting a predefined streamflow threshold, following an approach similar to Mason (1982) and Charpentier-Noyer et al. (2022). A contingency table (Table 2) is used, and four scores are considered (CAWCR 2015): critical success index (CSI), probability of detection (POD), success ratio (SR), and bias score (BIAS). Analysis is performed on the blending-driven method, where the sample of data is larger, and data from both periods are considered together. The chosen threshold is the 70th percentile of each benchmark streamflow time series, allowing for the selection of the main occurring peak events.
Contingency table, where Qb is the benchmark, Qf is the forecast (done with a certain lead time: 1, 2, …, hours), and Qt is a threshold. As in Charpentier-Noyer et al. (2022) and Godet et al. (2023), a tolerance window is considered in order to calculate a hit, miss, and false alarm. It is set to 3 h (half of the rainfall forecast time window), and this is considered by the authors a reasonable balance between excessive tolerance and the accounting of possible variability of the results from various subsequent predictions.
4. Results
a. Blending system (with 1-h updating)
The first verification is conducted exclusively using the blending system only with both assimilation methods in WRF. This method allows for updates every hour, resulting in a large sample of forecasts that contributes to robust statistical analysis.
To assess the performance, we constructed graphs depicting the forecast lead time on the x axis and the scores’ boxplot on the y axis. The boxplots present the interquartile range (25th and 75th percentiles) while the whiskers provide insight into the variability of the results relative to the median, and the mean is represented by the dot. These visualizations aid in understanding the distribution and variability of the scores across different forecast lead times.
The predicted rain impact on streamflow varies according to the basin size, which serves as an input integrator. Table 3 provides the number of samples used to construct the graphs for each basin class, categorized by the two periods. The number of samples depends on the sections involved and on the number of forecasts available.
Number of forecast samples used for building the graphs in Figs. 5–7 for the blending analysis.
Figure 5 shows the KGE for the two periods, October and November 2019. As suggested by Knoben et al. (2019), a KGE value of
KGE scores as a forecast lead time function for the blending method for the October period. The x axis reports lead time, and the y axis reports the boxplot of score values and the average value (dots). Rad indicates the radar data assimilation, while Rad+Light is the assimilation of both radar and lightning data. The “A” is the drainage area of the basins. The light gray area indicates that from the seventh hour rainfall input is not available and only the effects of hydrological model routing are shown.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
Both periods exhibit satisfactory KGE values, with the KGE remaining largely over 0 in both assimilation configurations within the 6-h rainfall forecast time window. This holds true for all the three classes of basins, although the performance tends to decrease at a slower rate with increasing lead time as the drainage area increases. Additionally, it is worth noting that events during the November period are generally more predictable compared to those in October.
During the October period, the KGE performances are quite similar for both assimilation systems. For drainage areas in the range 50–150 km2, Rad performs better for lead times between 3 and 6 h, while for areas larger than 150 km2, KGE is slightly better when Rad+Light is considered. In the November period, the Rad+Light configuration generally performs better for lead times ranging from 4 to 8 h. This variation in performance could be attributed to the fact that, depending on the affected area and lead time, the rainfall field produced with the Rad configuration may be better than Rad+Light in reproducing observed rainfall, and vice versa.
In both periods, there are no noticeable differences in the first 2 h of forecast, where the combination of nowcasting and routing of observed rainfall plays a significant role.
The relative errors (RelErr) are shown in Fig. 6, with values ranging between −0.15 and 0.15 for lead times up to 6 h. Beyond this limit, the errors rapidly decrease below −0.15 because the rainfall input is no longer available, and streamflow is solely influenced by the rainfall that occurred or is forecasted in the previous hours.
RelErr score as a forecast lead time function for the blending method.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
During the October period, the RelErr values are really close to 0 up to lead times of 3–5 h, with a small overall underestimation. The absolute errors (whiskers) are lower than 0.1 indicating relatively accurate streamflow predictions. In contrast, during the November period, the system tends to slightly overestimate the streamflow up to lead times of 5–6 h. The variability performance is smaller, particularly for drainage areas larger than 50 km2 where the interquartile range is confined between −0.05 and 0.05.
Figure 7 displays the correlation coefficient (Corr) graphs. The results indicate that the Rad+Light configuration generally performs better in terms of correlation coefficient for both the October and November periods. However, there are a few lead times during the October period where the Rad configuration outperforms the Rad+Light configuration, specifically for basins with a drainage area ranging from 50 to 150 km2. Overall, regardless of the configuration and period, the correlation coefficient remains relatively high, with interquartile ranges larger than 0.4–0.5. This suggests a strong relationship between the predicted and observed streamflow values.
Correlation coefficient ρ score as a forecast lead time function for the blending method.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
Figure 8 displays the comparison in terms of the selected categorial statistics as a function of lead time. The predictions are nearly to be unbiased for all the basin classes, with a slight tendency toward underestimation until 6–7 h, and for larger lead times, underestimation increases as the input becomes zero rain. CSI and SR for class 1 and class 2 are better for the Rad+Light configuration for lead times in the range of 3–7 h. This observation is consistent with a better localization of intense storms driven by lightning assimilation. However, for larger basins, in class 3 the catchment integration effect leads to really similar results for both configurations. The POD shows similar information to CSI and SR but with fewer differences in class 2. The BIAS lines for Rad and Rad+Light configurations for all the basin classes, indicating no significant differences arises between the two configurations in the terms of this metric.
Critical success index (CSI), BIAS score, success ratio (SR), and probability of detection (POD) as a function of forecast lead time. Results are shown for each basin class. Rad indicates the radar data assimilation, while Rad+Light is the assimilation of both radar and lightning data.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
b. Comparing blending and SWING systems (with 3-h updating)
In this section, we compare the results of the two rainfall forecast approaches, blending and SWING, with the two assimilation configurations. It is important to note that the number of forecasts available for comparison is lower in this section compared to the previous section. This is because we can compare results only when both SWING and blending are available, which occurs every 3 h. As a result, the statistical calculations for the blending approach are affected by this undersampling, leading to some changes in the results compared to when an update frequency of 1 h is considered. Table 4 provides the number of samples used for each basin class and for each to build the graphs.
Number of forecast samples used for building the graphs in Figs. 9–14 for the blending–SWING comparison.
Once again, for all four cases, KGE values remain larger than 0.0 for up to 6–7 h of lead time, which can be considered a good result. In Figs. 9 and 10, it can be observed that during the October period SWING seems to perform slightly better with very similar results using the Rad or Rad+Light assimilation. Blending is weakly better only for some lead times for basins with a drainage area in the range of 50–150 km2. On the other hand, for the November period, blending performs slightly better than SWING. However, it is not as evident as in previous section that the Rad+Light assimilation is better than Rad. This may be attributed to the undersampling mentioned earlier. As a result, we emphasize the importance of giving more weight to the results described in section 4a when making a decision regarding the use of either the Rad or Rad+Light configuration in a hypothetical operational framework.
Comparison between the blending and SWING methods. KGE scores as a forecast lead time function for the October period.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
As in Fig. 9, but for the November period.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
Furthermore, when considering RelErr (Figs. 11 and 12), it can be observed that SWING performs better in the October period reducing underestimations. On the other hand, blending works better in the November period reducing SWING overestimations, particularly when Rad+Light configuration is used. Presumably the different approaches used in model–observation comparison and dynamical parameterization of the equations in blending and SWING result in different rainfall forecasts. Once again, the analysis reveals some contrasts compared to considering the complete blending results with a 1-h update. In the latter case, the differences between using the Rad+Light or Rad configurations are minimal in terms of RelErr.
Comparison between the blending and SWING methods. RelErr score as a forecast lead time function for the October period.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
As in Fig. 11, but for the November period.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
The correlation coefficient demonstrates a very high correlation between the benchmark and forecast up to 6–7 h of lead time (Figs. 13 and 14), especially for basins with drainage area larger than 150 km2, which are less sensitive to small errors in precipitation location. However, there are inconsistencies compared to the results found for the 1-h update in blending approach, as an example for basins in class of drainage area between 50 and 150 km2. In the 1-h update analysis, the Rad+Light configuration performs better than the Rad configuration for lead times of 2–5 h. However, in the comparison between the SWING and blending approaches, the analysis of results leads to contrary considerations. As previously mentioned, these discrepancies can be attributed to the different samples used for the analysis.
Comparison between the blending and SWING methods. Corr score as a forecast lead time function for the October period.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
As in Fig. 13, but for the November period.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
As a general comment, it can be stated that none of the four configurations considered is definitively better than the others. The results indicate that each of the configurations can be the best option, depending on the specific period (October or November), the range of drainage areas, and the lead time. Considering the overall results, it can be concluded that the use of the Rad+Light assimilation configuration generally improves the results for both the blending and SWING approaches.
c. Analysis of subevents during October period
In this subsection, the October period forecasts are divided into three main subperiods just to examine how the system performs under different conditions. The objective is to gain an understanding of how the combination of extrapolation algorithm and NWP can behave for different kinds of events in terms of hydrological predictability. The October period is chosen since it includes subevents with different characteristics (stratiform and convective cases), and moreover during 14–15 and 21–22 October floods with lot of damages occurred. The analysis is presented using all the hourly forecasts of the blending configuration and comparing the two methods when both are available, as done in previous sections. For simplicity, results are presented in terms of KGE and RelErr only.
The three October subperiods have the following main characteristics:
-
14–15 October: This subperiod is characterized by a series of intense events that occurred on small areas with high persistence. The main event is very localized in time and space in the center of the study area, occurring on the 14 October evening.
-
16–20 October: This subperiod experienced widespread precipitation affecting large areas of the study domain.
-
21–22 October: The final subperiod is marked by a very intense longer-lasting persistent event that primarily affected a group of basins in the central part of the study area.
Table 5 reports the number of forecast samples used to carry out the analysis for each subperiod.
Number of forecast samples used for building the graphs in Figs. 15–17 for the October subperiods analysis.
Figures 15–17 present the KGE and RelErr separately for the three subperiods. The upper panels depict the analysis using blending only (update 1 h), while the bottom panels display the blending–SWING comparison (update 3 h). In the central days (16–20 October, Fig. 16) the predictability is higher, resulting in a weak decrease in performance with lead time.
KGE and RelErr as lead time function for the 14–15 Oct subperiod of the October use case. (a),(c) The blending method and hourly update. (b),(d) Both the blending and SWING methods are available (update: 3 h).
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
KGE and RelErr as lead time function for the 16–20 Oct subperiod of the October use case. (a),(c) The blending method and hourly update. (b),(d) Both the blending and SWING methods are available (update: 3 h).
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
KGE and RelErr as lead time function for the 21–22 Oct subperiod of the October use case. (a),(c) The blending method and hourly update. (b),(d) Both the blending and SWING methods are available (update: 3 h).
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0125.1
However, the other two subperiods (Figs. 15a,c and 17a,c) are characterized by lower predictability. Specifically, during the 21–22 October, there is a variability in the interquartile range, with mean and median values lower than 0.9 even for the first 2 h of forecast. These results are supported by the RelErr trends, where Fig. 17c indicates an overestimation in terms of RelErr during the first 2 h of forecast, consistent with KGE values. This is probably due to an overestimation of the rainfall extrapolated by PhaSt in the first hour at least in some of the involved basins. These findings align with the characteristic of the rainfall subperiods. Small and intense rainfall structures are more unpredictable, especially in terms of exact location and intensity (Molini et al. 2009). Similar occurrences have been observed when narrow but larger rainfall structures (length of 60–90 km and width 15–20 km) impact the area (Rebora et al. 2013; Silvestro et al. 2016) as seen during 21 October. Regarding the two different assimilation configurations (Rad and Rad+Light), they exhibit similar performance overall. However, Rad+Light shows slightly better KGE for the subperiods 14–15 and 21–22 October (Figs. 15a,b and 17a,b). This improvement is particularly noticeable for lead times between 5 and 8 h. It is worth noting, when considering RelErr (Fig. 17c), that during the 21–22 October subperiod, the Rad configuration leads to a larger overestimation for lead times shorter than 4 h and drainage areas smaller than 150 km2. However, it also brings RelErr closer to 0 for lead times between 5 and 8 h.
When comparing the blending and SWING methods, it is important to consider that the 3-h update interval can further affect the results due to the splitting of the entire period into shorter subperiods (Figs. 15–17, panels b and d). In terms of KGE, the results of the two methods are similar, but the use of SWING indicates a higher predictability of the 16–20 October subperiod. For certain lead times (4–7 h) in the 14–15 and 16–20 October subperiods, slightly better values are observed for SWING, particularly when considering basins with area larger than 50 km2. Noteworthy differences in KGE values for the 14–15 October subperiod emerge depending on the assimilation configurations. For basins smaller than 50 km2, Rad+Light improves performance up to 7–8 h of lead time for both blending and SWING methods (Figs. 15a,b).
SWING demonstrates slightly better performances in terms of RelErr for lead times between 5 and 7 h in the 14–15 October subperiod when using the Rad+Light configuration. Similarly, in the 16–20 October subperiod SWING performs slightly better with the Rad configuration. However, during the 21–22 October subperiod, the best option in terms of RelErr appears to be the blending method with the Rad+Light configuration for low lead times (up to 3–4 h), and the blending method with the Rad configuration for lead times up to 7–8 h.
5. Conclusions
The presented study explores the capabilities of two short-term flood forecast systems in predicting streamflow during heavy rainfall events.
Both systems incorporate two key components: an extrapolation nowcasting system and an NWP model. The two approaches leverage observations and information exchange in different ways. Specifically, a 3DVAR assimilation method is employed to assimilate radar reflectivity, while a nudging technique is used for lightning assimilation in NWP. Additionally, the precipitation volume trend of the latter is used to overcome limitations associated with the extrapolation nowcasting method. The method called blending utilizes rainfall observations to dynamically parameterize a linear combination function between nowcasting and the most updated NWP run. On the other hand, the method named SWING, in addition to using a linear combination system between nowcasting and NWP similar to the blending method, leverages the overlapping time windows of the three recent NWP run weighting each forecast considering its behavior in the previous 3 h. The flood forecast is then carried out with a distributed hydrological model for both methods.
Since two short-term forecasts approaches and two possible assimilation configurations were considered, four system settings were tested in a hindcast experiment during two rainy periods. The following considerations can be made.
All four settings provide good results for streamflow prediction, with performance being generally satisfactory up to 5–8 h of lead time. The KGE consistently remains larger than 0 and greater than 0.3 in the first 4–6 h of the forecast. The RelErr remains within the range between −0.15 and 0.15 and averages close to 0.0, while the Corr values are high.
It is not possible to definitely define a configuration that is always better than the others. When considering different lead times, different analysis periods, and basin classes, results show that one of the four settings can be better in respect to the others in specific circumstances.
Without contradicting the previous statements, and taking an overall look at the results, it can be stated that using both radar data and lightning in the assimilation process provides an average added value to the performance. However, under certain conditions it appears that using radar data alone performs better, especially for less intense events.
In terms of performance, blending and SWING seem to closely compete with each other. In some cases, SWING emerges as the best option highlighting the benefit of considering various NWP runs. However, in other cases, the blending approach performs better.
Considering the three different basin classes, an expected trend emerged, although not in a stable and definite manner, as shown in Heuvelink et al. (2020). When basin dimension increases, performance tends to increase due to the catchment’s integration effect of the catchments themselves. However, when comparing different metrics, the differences for the three classes are sometimes negligible, and results for the October 2019 period (lower predictability) differ from the November 2019 period (higher predictability).
Blending self-standing analysis reveals partially different results compared to the blending–SWING comparison. The first investigation involved a larger number of available forecasts (forecast frequency 1 h instead of 3 h) and can be considered more robust. By examining the results we gain insight into how the sample size can impact the final scores values even when analyzing the same period.
A final qualitative consideration can be made from a possible operational perspective. The goal would be to reduce the computational efforts while effectively managing the uncertainties and leveraging the different capabilities of the four analyzed configurations. In this context, using both blending and SWING configurations with assimilation of radar and lightning data could be a favorable option.
The blending algorithm allows for a more frequent update of the streamflow forecast (e.g., 1 h) including the very most recent data; SWING drives a less frequent update (3 h) but exploits the information content of three different runs of the NWP, and this should reduce the false alarms as previously shown in Lagasio et al. (2022a). Both algorithms benefit from assimilation (Poletti et al. 2019; Lagasio et al. 2019, 2022a), especially when radar and lightning data are assimilated, which seems the best solution on average. It can be thus stated that using two streamflow forecast chains (with a mini ensemble of two members when both are available) is the best solution to improve streamflow predictions and exploiting as best as possible all the data and modeling information.
The potential flexibility of the framework is indeed high. Both the blending and SWING strategies can be applied with nowcasting methods and meteorological models that share characteristics similar to those used in the presented analysis. Analogous considerations can be made regarding the hydrological model. Additionally, the SWING method can be utilized in an ensemble mode with other meteorological forecasts, adopting a multimodel approach. This enables the integration of diverse meteorological predictions, enhancing the forecasting capabilities and potentially providing more reliable and comprehensive flood predictions. The adaptable nature of the framework makes it suitable for various settings and opens possibilities for future improvements and extensions.
Indeed, further enhancements to the present work could involve the implementation of a stochastic version of the nowcasting extrapolation method. However, it is essential to consider that this would lead to an increase in computational requirements. Additionally, it would be interesting to test the system on events that involve other geographical areas covered by the hydrometeorological system described in section 2c. This would allow for a broader assessment of the system’s performance under different weather conditions and geographical characteristics.
Indeed, another avenue for improvement could involve modifying the extrapolation method itself to address its limitations. For instance, investigating the feasibility of introducing an orographic feature contribution to the nowcasting process could potentially enhance the accuracy of the forecasts, especially in regions with complex terrain. Additionally, further enhancements in the assimilation process within NWP could be explored by incorporating other types of observations.
In conclusion, the presented study demonstrates the effectiveness of the two short-term flood forecast systems and highlights opportunities for future developments to enhance their capabilities and applicability in operational settings.
Acknowledgements.
This research was co-funded by the Italian Department of Civil Protection, Presidency of the Council of Ministers, through a convention between Department of Civil Protection (CUP B57G22000400001), by the PNRR RAISE (Robotics and AI for Socio Economic Empowerment, CUP B57G22000780006) and CIMA Research foundation.
Data availability statement.
The authors confirm that the data supporting the findings of this study are available within the article. Derived data supporting the findings of this study are available on request from the corresponding author.
REFERENCES
Atencia, A., A. Kann, Y. Wang, and F. Meier, 2020: Localized variational blending for nowcasting purposes. Meteor. Z., 29, 247–261, https://doi.org/10.1127/metz/2020/1003.
Barker, D., and Coauthors, 2012: The Weather Research and Forecasting Model’s Community Variational/Ensemble Data Assimilation System: WRFDA. Bull. Amer. Meteor. Soc., 93, 831–843, https://doi.org/10.1175/BAMS-D-11-00167.1.
Beljaars, A. C. M., 1995: The parametrization of surface fluxes in large-scale models under free convection. Quart. J. Roy. Meteor. Soc., 121, 255–270, https://doi.org/10.1002/qj.49712152203.
Berenguer, M., C. Corral, R. Sánchez-Diezma, and D. Sempere-Torres, 2005: Hydrological validation of a radar-based nowcasting technique. J. Hydrometeor., 6, 532–549, https://doi.org/10.1175/JHM433.1.
Borga, M., 2002: Accuracy of radar rainfall estimates for streamflow simulation. J. Hydrol., 267, 26–39, https://doi.org/10.1016/S0022-1694(02)00137-3.
Bowler, N. E., C. E. Pierce, and A. W. Seed, 2006: STEPS: A probabilistic precipitation forecasting scheme which merges and extrapolation nowcast with downscaled NWP. Quart. J. Roy. Meteor. Soc., 132, 2127–2155, https://doi.org/10.1256/qj.04.100.
Brandolini, P., and P. Sbardella, 2001: Caratterizzazione del reticolo idrografico del territorio comunale di Genova a seguito delle modificazioni antropiche. Boll. Soc. Geogr. It., 12, 199–218.
Brandolini, P., and M. Spotorno, 1997: Distribuzione della popolazione per fasce altimetriche. La Popolazione della Liguria dal 1971 al 1991, D. Ruocco, ed., Istituto di Geografia dell’Universita di Genova, 67–91.
Braud, I., M. Borga, J. Gourley, M. Hürlimann, M. Zappa, and F. Gallart, 2016: Flash floods, hydro-geomorphic response and risk management. J. Hydrol., 541, 1–5, https://doi.org/10.1016/j.jhydrol.2016.08.005.
Bruno, G., F. Pignone, F. Silvestro, S. Gabellani, F. Schiavi, N. Rebora, P. Giordano, and M. Falzacappa, 2021: Performing hydrological monitoring at a national scale by exploiting rain-gauge and radar networks: The Italian case. Atmosphere, 12, 771, https://doi.org/10.3390/atmos12060771.
Buzzi, A., S. Davolio, P. Malguzzi, O. Drofa, and D. Mastrangelo, 2014: Heavy rainfall episodes over Liguria in autumn 2011: Numerical forecasting experiments. Nat. Hazards Earth Syst. Sci., 14, 1325–1340, https://doi.org/10.5194/nhess-14-1325-2014.
CAWCR, 2015: Forecast verification. Accessed 27 July 2023, https://www.cawcr.gov.au/projects/verification/.
Charpentier-Noyer, M., and Coauthors, 2022: A methodological framework for the evaluation of short-range flash-flood hydrometeorological forecasts at the event scale. Nat. Hazards Earth Syst. Sci., 23, 2001–2029, https://doi.org/10.5194/nhess-23-2001-2023.
Davolio, S., F. Silvestro, and P. Malguzzi, 2015: Effects of increasing horizontal resolution in a convection permitting model on flood forecasting: The 2011 dramatic events in Liguria (Italy). J. Hydrometeor., 16, 1843–1856, https://doi.org/10.1175/JHM-D-14-0094.1.
Davolio, S., F. Silvestro, and T. Gastaldo, 2017: Impact of rainfall assimilation on high-resolution hydro-meteorological forecasts over Liguria (Italy). J. Hydrometeor., 18, 2659–2680, https://doi.org/10.1175/JHM-D-17-0073.1.
De Leonibus, L., D. Biron, P. Laquale, F. Zauli, and D. Melfi, 2008: Rainfall field reconstruction over Italy through LAMPINET lightning data. 20th Int. Lightning Detection Conf., Tucson, AZ, Vaisala, 1–6, https://www.vaisala.com/sites/default/files/documents/Rainfall%20Field%20Reconstruction%20Over%20Italy%20Through%20LAMPINET%20Lightning%20Data.PDF.
De Leonibus, L., D. Biron, M. Sist, D. Labate, F. Zauli, and D. Melfi, 2010: Wind intensity reconstruction over Italy through LAMPINET lightning data. Third Int. Lightning Meteorology Conf., Orlando, FL, Vaisala, 1–6, https://www.vaisala.com/sites/default/files/documents/2.De%20Leonibus%2C%20Biron%2C%20Sist_0.pdf.
Delrieu, G., and Coauthors, 2005: The catastrophic flash-flood event of 8–9 September 2002 in the Gard Region, France: A first case study for the Cévennes–Vivarais Mediterranean Hydrometeorological Observatory. J. Hydrometeor., 6, 34–52, https://doi.org/10.1175/JHM-400.1.
DPC, 2022: Monitoraggio e sorveglianza. Accessed 19 December 2022, https://rischi.protezionecivile.gov.it/it/meteo-idro/attivita/centro-funzionale-centrale/monitoraggio-e-sorveglianza.
Dyer, A. J., and B. B. Hicks, 1970: Flux-gradient relationships in the constant flux layer. Quart. J. Roy. Meteor. Soc., 96, 715–721, https://doi.org/10.1002/qj.49709641012.
Edouard, S., B. Vincendon, and V. Ducrocq, 2018: Ensemble-based flash-flood modelling: Taking into account hydrodynamic parameters and initial soil moisture uncertainties. J. Hydrol., 560, 480–494, https://doi.org/10.1016/j.jhydrol.2017.04.048.
Faccini, F., F. Luino, G. Paliaga, A. Roccati, and L. Turconi, 2021: Flash flood events along the west Mediterranean coasts: Inundations of urbanized areas conditioned by anthropic impacts. Land, 10, 620, https://doi.org/10.3390/land10060620.
Fierro, A. O., E. R. Mansell, C. L. Ziegler, and D. R. MacGorman, 2012: Application of a lightning data assimilation technique in the WRFARW model at cloud-resolving scales for the tornado outbreak of 24 May 2011. Mon. Wea. Rev., 140, 2609–2627, https://doi.org/10.1175/MWR-D-11-00299.1.
Fiori, E., L. Ferraris, L. Molini, F. Siccardi, D. Kranzlmueller, and A. Parodi, 2017: Triggering and evolution of a deep convective system in the Mediterranean Sea: Modelling and observations at a very fine scale. Quart. J. Roy. Meteor. Soc., 143, 927–941, https://doi.org/10.1002/qj.2977.
Foresti, L., M. Reyniers, A. Seed, and L. Delobbe, 2016: Development and verification of a real-time stochastic precipitation nowcasting system for urban hydrology in Belgium. Hydrol. Earth Syst. Sci., 20, 505–527, https://doi.org/10.5194/hess-20-505-2016.
Franch, G., D. Nerini, M. Pendesini, L. Coviello, G. Jurman, and C. Furlanello, 2020: Precipitation nowcasting with orographic enhanced stacked generalization: Improving deep learning predictions on extreme events. Atmosphere, 11, 267, https://doi.org/10.3390/atmos11030267.
Gastaldo, T., V. Poli, T. Paccagnella, and P. P. Alberoni, 2022: Direct assimilation of radar reflectivity volumes in COSMO–LETKF. Precipitation Science, S. Michaelides, Ed., Elsevier, 757–786, https://doi.org/10.1016/B978-0-12-822973-6.00024-X.
Godet, J., O. Payrastre, P. Javelle, and F. Bouttier, 2023: Assessing the ability of a new seamless short-range ensemble rainfall product to anticipate flash floods in the French Mediterranean area. Nat. Hazards Earth Syst. Sci., 23, 3355–3377, https://doi.org/10.5194/nhess-23-3355-2023.
Guastavino, S., and Coauthors, 2022: Prediction of severe thunderstorm events with ensemble deep learning and radar data. Sci. Rep., 12, 20049, https://doi.org/10.1038/s41598-022-23306-6.
Han, J., and H.-L. Pan, 2011: Revision of convection and vertical diffusion schemes in the NCEP Global Forecast System. Wea. Forecasting, 26, 520–533, https://doi.org/10.1175/WAF-D-10-05038.1.
Heuvelink, D., M. Berenguer, C. C. Brauer, and R. Uijlenho, 2020: Hydrological application of radar rainfall nowcasting in the Netherlands. Environ. Int., 136, 105431, https://doi.org/10.1016/j.envint.2019.105431.
Hong, S. Y., and J. O. J. Lim, 2006: The WRF single-moment 6-class microphysics scheme (WSM6). Korean Meteor. Soc., 42, 129–151.
Hong, S.-Y., Y. Noh, and J. Dudhia, 2006: A new vertical diffusion package with an explicit treatment of entrainment processes. Mon. Wea. Rev., 134, 2318–2341, https://doi.org/10.1175/MWR3199.1.
Hwang, Y., A. J. Clark, V. Lakshmanan, and S. E. Koch, 2015: Improved nowcasts by blending extrapolation and model forecasts. Wea. Forecasting, 30, 1201–1217, https://doi.org/10.1175/WAF-D-15-0057.1.
Iacono, M. J., J. S. Delamere, E. J. Mlawer, M. W. Shephard, S. A. Clough, and W. D. Collins, 2008: Radiative forcing by long-lived greenhouse gases: Calculations with the AER radiative transfer models. J. Geophys. Res., 113, D13103, https://doi.org/10.1029/2008JD009944.
Ide, K., P. Courtier, M. Ghil, and A. C. Lorenc, 1997: Unified notation for data assimilation: Operational, sequential and variational. J. Meteor. Soc. Japan, 75, 181–189, https://doi.org/10.2151/jmsj1965.75.1B_181.
Imhoff, R. O., C. C. Brauer, A. Overeem, A. H. Weerts, and R. Uijlenhoet, 2020: Spatial and temporal evaluation of radar rainfall nowcasting techniques on 1,533 events. Water Resour. Res., 56, e2019WR026723, https://doi.org/10.1029/2019WR026723.
Imhoff, R. O., C. C. Brauer, K. J. van Heeringen, R. Uijlenhoet, and A. H. Weerts, 2022: Large‐sample evaluation of radar rainfall nowcasting for flood early warning. Water Resour. Res., 58, e2021WR031591, https://doi.org/10.1029/2021WR031591.
Knoben, W. J. M., J. E. Freer, and R. A. Woods, 2019: Technical note: Inherent benchmark or not? Comparing Nash–Sutcliffe and Kling–Gupta efficiency scores. Hydrol. Earth Syst. Sci., 23, 4323–4331, https://doi.org/10.5194/hess-23-4323-2019.
Lagasio, M., A. Parodi, R. Procopio, F. Rachidi, and E. Fiori, 2017: Lightning potential index performances in multimicrophysical cloud-resolving simulations of a back-building mesoscale convective system: The Genoa 2014 event. J. Geophys. Res. Atmos., 122, 4238–4257, https://doi.org/10.1002/2016JD026115.
Lagasio, M., F. Silvestro, L. Campo, and A. Parodi, 2019: Predictive capability of a high-resolution hydrometeorological forecasting framework coupling WRF cycling 3DVAR and continuum. J. Hydrometeor., 20, 1307–1337, https://doi.org/10.1175/JHM-D-18-0219.1.
Lagasio, M., and Coauthors, 2022a: SWING, the score-weighted improved nowcasting algorithm: Description and application. Water, 14, 2131, https://doi.org/10.3390/w14132131.
Lagasio, M., and Coauthors, 2022b: A complete meteo/hydro/hydraulic chain application to support early warning and monitoring systems: The Apollo medicane use case. Remote Sens., 14, 6348, https://doi.org/10.3390/rs14246348.
Liguori, S., M. A. Rico-Ramirez, A. N. A. Schellart, and A. J. Saul, 2012: Using probabilistic radar rainfall nowcasts and NWP forecasts for flow prediction in urban catchments. Atmos. Res., 103, 80–95, https://doi.org/10.1016/j.atmosres.2011.05.004.
Lovat, A., B. Vincendon, and V. Ducrocq, 2022: Hydrometeorological evaluation of two nowcasting systems for Mediterranean heavy precipitation events with operational considerations. Hydrol. Earth Syst. Sci., 26, 2697–2714, https://doi.org/10.5194/hess-26-2697-2022.
Mason, I., 1982: A model for assessment of weather forecasts. Aust. Meteor. Mag., 30, 291–303.
Mazzarella, V., and Coauthors, 2022: Is an NWP-based nowcasting system suitable for aviation operations? Remote Sens., 14, 4440, https://doi.org/10.3390/rs14184440.
Metta, S., J. von Hardernberg, L. Ferraris, N. Rebora, and A. Provenzale, 2009: Precipitation nowcasting by a spectral-based nonlinear stochastic model. J. Hydrometeor., 10, 1285–1297, https://doi.org/10.1175/2009JHM1120.1.
Molini, L., A. Parodi, and F. Siccardi, 2009: Dealing with uncertainty: An analysis of the severe weather events over Italy in 2006. Nat. Hazards Earth Syst. Sci., 9, 1775–1786, https://doi.org/10.5194/nhess-9-1775-2009.
Pagliara, P., A. Corina, A. Burastero, P. Campanella, L. Ferraris, M. Morando, N. Rebora, and C. Versace, 2011: Dewetra, coping with emergencies. Proc. Eighth Int. ISCRAM Conf., Lisbon, Portugal, ISCRAM, 1–5, http://idl.iscram.org/files/pagliara/2011/827_Pagliara_etal2011.pdf.
Panziera, L., U. Germann, M. Gabella, and P. V. Mandapaka, 2011: NORA–Nowcasting of orographic rainfall by means of analogues. Quart. J. Roy. Meteor. Soc., 137, 2106–2123, https://doi.org/10.1002/qj.878.
Parodi, A., M. Lagasio, A. N. Meroni, F. Pignone, F. Silvestro, and L. Ferraris, 2020: A hindcast study of the Piedmont 1994 flood: The CIMA research foundation hydro-meteorological forecasting chain. Bull. Atmos. Sci. Technol., 1, 297–318, https://doi.org/10.1007/s42865-020-00023-4.
Paulson, C. A., 1970: The mathematical representation of wind speed and temperature profiles in the unstable atmospheric surface layer. J. Appl. Meteor., 9, 857–861, https://doi.org/10.1175/1520-0450(1970)009<0857:TMROWS>2.0.CO;2.
Pirone, D., L. Cimorelli, G. D. Giudice, and D. Pianese, 2023: Short-term rainfall forecasting using cumulative precipitation fields from station data: A probabilistic machine learning approach. J. Hydrol., 617, 128949, https://doi.org/10.1016/j.jhydrol.2022.128949.
Poletti, M. L., F. Silvestro, S. Davolio, F. Pignone, and N. Rebora, 2019: Using nowcasting technique and data assimilation in a meteorological model to improve very short range hydrological forecasts. Hydrol. Earth Syst. Sci., 23, 3823–3841, https://doi.org/10.5194/hess-23-3823-2019.
Ramos Filho, G. M., V. H. R. Coelho, E. da Silva Freitas, Y. Xuan, and C. da Neves Almeida, 2021: An improved rainfall-threshold approach for robust prediction and warning of flood and flash flood hazards. Nat. Hazards, 105, 2409–2429, https://doi.org/10.1007/s11069-020-04405-x.
Rebora, N., and Coauthors, 2013: Extreme rainfall in the Mediterranean: What can we learn from observations? J. Hydrometeor., 14, 906–922, https://doi.org/10.1175/JHM-D-12-083.1.
Roccati, A., F. Luino, L. Turconi, P. Piana, C. Watkins, and F. Faccini, 2018: Historical geomorphological research of a Ligurian coastal floodplain (Italy) and its value for management of flood risk and environmental sustainability. Sustainability, 10, 3727, https://doi.org/10.3390/su10103727.
Sideris, I. V., L. Foresti, D. Nerini, and U. Germann, 2020: NowPrecip: Localized precipitation nowcasting in the complex terrain of Switzerland. Quart. J. Roy. Meteor. Soc., 146, 1768–1800, https://doi.org/10.1002/qj.3766.
Silvestro, F., and N. Rebora, 2012: Operational verification of a framework for the probabilistic nowcasting of river discharge in small and medium size basins. Nat. Hazards Earth Syst. Sci., 12, 763–776, https://doi.org/10.5194/nhess-12-763-2012.
Silvestro, F., S. Gabellani, F. Delogu, R. Rudari, and G. Boni, 2013: Exploiting remote sensing land surface temperature in distributed hydrological modelling: The example of the Continuum model. Hydrol. Earth Syst. Sci., 17, 39–62, https://doi.org/10.5194/hess-17-39-2013.
Silvestro, F., S. Gabellani, F. Delogu, R. Rudari, P. Laiolo, and G. Boni, 2015: Uncertainty reduction and parameter estimation of a distributed hydrological model with ground and remote-sensing data. Hydrol. Earth Syst. Sci., 19, 1727–1751, https://doi.org/10.5194/hess-19-1727-2015.
Silvestro, F., and Coauthors, 2016: What if the 25 October 2011 event that struck Cinque Terre (Liguria) had happened in Genoa, Italy? Flooding scenarios, hazard mapping and damage estimation. Nat. Hazards Earth Syst. Sci., 16, 1737–1753, https://doi.org/10.5194/nhess-16-1737-2016.
Silvestro, F., A. Parodi, L. Campo, and L. Ferraris, 2018: Analysis of the streamflow extremes and long-term water balance in the Liguria region of Italy using a cloud-permitting grid spacing reanalysis dataset. Hydrol. Earth Syst. Sci., 22, 5403–5426, https://doi.org/10.5194/hess-22-5403-2018.
Silvestro, F., G. Ercolani, S. Gabellani, P. Giordano, and M. Falzacappa, 2021: Improving real-time operational streamflow simulations using discharge data to update state variables of a distributed hydrological model. Hydrol. Res., 52, 1239–1260, https://doi.org/10.2166/nh.2021.162.
Sinclair, S., and G. Pegram, 2005: Combining radar and rain gauge rainfall estimates using conditional merging. Atmos. Sci. Lett., 6, 19–22, https://doi.org/10.1002/asl.85.
Skamarock, W. C., and Coauthors, 2008: A description of the Advanced Research WRF version 3. NCAR Tech. Note NCAR/TN-475+STR, 113 pp., https://doi.org/10.5065/D68S4MVH.
Smirnova, T. G., J. M. Brown, and S. G. Benjamin, 1997: Performance of different soil model configurations in simulating ground surface temperature and surface fluxes. Mon. Wea. Rev., 125, 1870–1884, https://doi.org/10.1175/1520-0493(1997)125<1870:PODSMC>2.0.CO;2.
Smirnova, T. G., J. M. Brown, S. G. Benjamin, and D. Kim, 2000: Parameterization of cold season processes in the MAPS land-surface scheme. J. Geophys. Res., 105, 4077–4086, https://doi.org/10.1029/1999JD901047.
Sofokleous, I., A. Bruggeman, C. Camera, and M. Eliades, 2022: Grid-based calibration of the WRF-hydro with Noah-MP model with improved groundwater and transpiration process equations. J. Hydrol., 617, 128991, https://doi.org/10.1016/j.jhydrol.2022.128991.
Vieux, B. E., and P. B. Bedient, 2004: Assessing urban hydrologic prediction accuracy through event reconstruction. J. Hydrol., 299, 217–236, https://doi.org/10.1016/S0022-1694(04)00366-X.
Vulpiani, G., M. Montopoli, L. D. Passeri, A. G. Gioia, P. Giordano, and F. S. Marzano, 2012: On the use of dual-polarized C-band radar for operational rainfall retrieval in mountainous areas. J. Appl. Meteor. Climatol., 51, 405–425, https://doi.org/10.1175/JAMC-D-10-05024.1.
Vulpiani, G., E. Guerriero, P. Giordano, M. Negri, and P. Pagliara, 2016: The Italian radar QPE: Description performance analysis and perspectives. Plinius Conference Abstracts, Vol. 15, Plinius15-69, https://meetingorganizer.copernicus.org/Plinius15/Plinius15-69.pdf.
Wang, F., G. Huang, G. Cheng, and Y. Li, 2021: Multi-level factorial analysis for ensemble data-driven hydrological prediction. Adv. Water Resour., 153, 103948, https://doi.org/10.1016/j.advwatres.2021.103948.
Wang, H., X.-Y. Huang, J. Sun, D. Xu, M. Zhang, S. Fan, and J. Zhong, 2014: Inhomogeneous background error modeling for WRF-Var using the NMC method. J. Appl. Meteor. Climatol., 53, 2287–2309, https://doi.org/10.1175/JAMC-D-13-0281.1.
Webb, E. K., 1970: Profile relationships: The log-linear range, and extension to strong stability. Quart. J. Roy. Meteor. Soc., 96, 67–90, https://doi.org/10.1002/qj.49709640708.
Zhang, Y., M. Long, K. Chen, L. Xing, R. Jin, M. I. Jordan, and J. Wang, 2023: Skilful nowcasting of extreme precipitation with NowcastNet. Nature, 619, 526–532, https://doi.org/10.1038/s41586-023-06184-4.
