Comment on “Can assimilation of crowdsourced data in hydrological modelling improve flood prediction?” by Mazzoleni et al. (2017)

2.1 The Bacchiglione catchment closed at Ponte degli Angeli (Vicenza)

Figure 1The catchment of the Bacchiglione River at Ponte degli Angeli, Vicenza (Italy).

Download

The catchment of the upper Bacchiglione River, closed at Ponte degli Angeli in the historical centre of Vicenza (Fig. 1), is located in the north of the Veneto Region, a plain that is fringed by the Alpine barrier at a distance of less than 100 km north of the Adriatic Sea (Barbi et al., 2012).

With regard to the precipitation climatology, the southern part of this plain is the drier one, with approximately 700–1000 mm of mean annual rainfall, whereas more than 2000 mm are measured close to the pre-Alpine chain due to the interaction of the southerly warm and humid currents coming from the Mediterranean Sea with the mountain barrier (Smith, 1979). A significant portion of the annual rainfall often concentrates into very short periods of time in the form of what often turns out to be an extreme event with deep convection playing a central role (Barbi et al., 2012; Rysman et al., 2016). As a consequence, severe flooding events have threatened agricultural and urban areas in recent years (e.g. Viero et al., 2013; Scorzini and Frank, 2017).

Due to the spatial and temporal variability of the rainfall fields, meteorological models are often unable to provide accurate and reliable quantitative precipitation estimates for the upper Bacchiglione catchment. An example of this inadequacy is given, for instance, by Fig. 13 in Mazzoleni et al. (2017).

The upper Veneto plain is a highly populated and urbanized area, with extremely complex drainage and irrigation networks that significantly affect both runoff production and propagation (Viero and Valipour, 2017). Within this plain, the Bacchiglione River and its tributaries are provided with relatively high levees (Viero et al., 2013), which prevent the exchange of water from inside to outside the riverbed (and vice versa) when the inner water levels are relatively high. As a consequence, the minor channel networks are not always allowed to deliver their drainage water towards the nearest tributary, i.e. the inflow points along the main river reaches change during a flood event depending on the instantaneous water level within the river. This occurrence modifies the network connectedness which, in turn, leads to different mechanisms of hydrologic response in the overall catchment.

Just upstream of the city of Vicenza, an area of up to 1 km² (the “Viale Diaz” floodplain, Fig. 1) is flooded when the Bacchiglione flow rate exceeds ∼ 160 m³ s⁻¹. Since about 2 × 10⁶ m³ of water can be temporarily stored in this area, a significant flood attenuation can be produced, particularly in the case of hydrographs with a steep rising limb (which is often the case due to the climatic regime and the catchment characteristics).

Moreover, the lower part of the Bacchiglione basin, north of Vicenza, includes a vast groundwater resurgence zone, in which it is difficult to assess both the actual contribution of resurgence to the Bacchiglione streamflow (up to ∼30 m³ s⁻¹) and the time-variable behaviour of soil moisture.

Clearly, such a system is highly non-linear. Nonetheless, significant parts of the Bacchiglione catchment are poorly monitored, and the remaining parts are completely unmonitored. The Leogra subcatchment (blue shaded area in Fig. 1) is provided with a pressure transducer for the measurement of water level at Torrebelvicino (Fig. 1). A rating curve derived from theoretical considerations is available for this cross section. However, the absence of instrumental measurements of flow discharge limits its reliability. The Leogra–Timonchio subcatchment (orange shaded area in Fig. 1) is monitored by an ultrasonic stage sensor located at Ponte Marchese, just upstream of the confluence with the Orolo River. Flow rate measurements at Ponte Marchese refer only to low hydraulic regimes, and show great variability due to the operation of a hydroelectric power plant located just downstream of Ponte Marchese. The Orolo River (green shaded area in Fig. 1), with a discharge capacity of more than one-third of the Bacchiglione at Ponte degli Angeli, is one of its major tributaries. Unfortunately, not only is the Orolo subcatchment completely uncovered by meteorological gauging stations, but no hydrometric gauging stations are present along its reach either. Similarly to the Orolo, the Astichello catchment (red shaded area in Fig. 1) is unmonitored and, due to backwater effects, significant areas adjacent to the main channel of the Astichello are flooded when water levels in the Bacchiglione are relatively high. Hence, the discharge that effectively flows from the Astichello into the Bacchiglione River may significantly decrease depending on the water stage within the main course of the Bacchiglione River.

Attention must be paid to the fact that the three major tributaries (Orolo, Timonchio, and Astichello) meet just upstream of the gauging station of Ponte degli Angeli (Fig. 1), making it difficult to correctly estimate the actual contribution of each single tributary to the total streamflow. By looking at the tree-like structure of the drainage network in an electrical analogy (Rodríguez-Iturbe and Rinaldo, 2001), the major tributaries of the Bacchiglione are in fact “conductors in parallel”.

Certainly, given the irregular topography of the catchments, the heterogeneity of the landscape, and the complexity of the hydraulic network, it can be stated that the Bacchiglione catchment is poorly monitored.

2.2 The semi-distributed model of the Bacchiglione catchment

In catchments like that of the Bacchiglione River, for all the reasons reported in the previous section, the accurate prediction of flood hydrographs with continuous time simulation is unquestionably a difficult task (Anquetin et al., 2010).

Mazzoleni et al. (2017) used an available semi-distributed hydrological model coupled with a Muskingum–Cunge scheme for flood propagation within the main river network, which was originally set up to forecast flood hydrographs of the Bacchiglione River at Ponte degli Angeli (Vicenza). Sensibly, the model was calibrated by minimizing the root mean square error between observed and simulated values of water discharge only at Ponte degli Angeli, which is the only hydrometric station provided with a reliable rating curve. The semi-distributed model, although explicitly representing the hydrological processes within the main subcatchments, has to be intended as a lumped model from a practical standpoint, since the discharge in Ponte degli Angeli is its only control point.

Therefore, regardless of the accuracy of streamflow prediction in Ponte degli Angeli, little can be said about the accuracy of the model in describing the internal states of the system, such as the streamflow along upstream tributaries. This limitation has to be ascribed to uncertainty in precipitation fields, to the paucity of (reliable) flow rate data upstream of Vicenza, and to inherent limitations of the model itself.

Indeed, it has to be remarked that the Muskingum–Cunge model for flood propagation used in Mazzoleni et al. (2017) considers rectangular river cross sections for the estimation of hydraulic radius, wave celerity, and other hydraulic variables (Todini, 2007). Accordingly, the effects exerted by the Viale Diaz floodplain, which acts as a sort of in-line natural flood control reservoir on flood propagation, can not be properly accounted for. This means that, if the flood hydrograph is correctly modelled at Ponte degli Angeli, it can not be correctly modelled upstream of the Viale Diaz floodplain (and vice versa).

2.3 The use of synthetic CSD in the Bacchiglione case study

In the Bacchiglione case study, Mazzoleni et al. (2017) calibrated the model using measured rainfall data to reproduce the streamflow hydrograph at the basin outlet well (call this post-event simulation “scenario 1”). Then they forced the model with predicted rainfall fields that were completely different from the actual storm event (“scenario 2”); in this case, the discharge simulated using forecasted input was very different from that obtained using recorded rainfall, with a significant time shift and errors in predicted discharge ranging between 25 and 50 % at the flood peak (and up to 90 % when considering synchronous data). In their “scenario 3”, similarly to the “observing system simulation experiment” (OSSE) approach, synthetic streamflow CSD extracted from “scenario 1” were assimilated into a new run using the same forcing as in “scenario 2”. Not surprisingly, the model performance in “scenario 3” was significantly better than in “scenario 2”, as the synthetic CSD they assimilated were representative of the model internal states in the best-fit scenario.

The authors argued that the synthetic CSD they used are realistic. For this condition to be met, given that these CSD are the results of the model itself, the model must well represent the physics of the real system (i.e. it must be calibrated or, at least, verified) at locations where CSD are first generated and then assimilated; this is a fundamental hypothesis behind the OSSE approach. The synthetic CSD used in Mazzoleni et al. (2017) for the Bacchiglione case study are drawn from the model internal states under best-fit conditions. Thus, when the model is forced with different (wrong) input data, their assimilation is expected to be as successful as possible in updating the model states toward the best-fit scenario. However, the accuracy of such synthetic CSD is questionable, since they do not refer to model control points (i.e. they are drawn from the semi-distributed model at locations where the model is neither calibrated nor verified), so nothing can actually be said about the model performance at these locations. In a sense, synthetic CSD used by Mazzoleni et al. (2017) are optimal (in view of assimilation performance) rather than realistic. Since real CSD are likely biased with respect to the synthetic CSD actually used, assimilation of real CSD can not be as effective as that performed in Mazzoleni et al. (2017).

From one point of view, it is possible that such an inconsistency could have led Mazzoleni et al. (2017) to overrate the importance of CSD, as they considered issues related to CSD precision, but not accuracy (Mazzoleni et al., 2017). Therefore, additional care must be taken in operational flood forecasting when assimilating CSD into (semi-)distributed hydrological models at locations other than model control points.