account for the intrinsic uncertainty contained into hydrological observations and
model structure, states and parameters. Two case studies, the Brue and Bacchiglione
catchments, are considered. Finally, the evaluation of the developed methods is
provided. This study demonstrates that networks of low-cost static and dynamic
social sensors can complement traditional networks of static physical sensors, for the
purpose of improving flood forecasting accuracy. This can be a potential application
of recent efforts to build citizen observatories of water, in which citizens not only can
play an active role in information capturing, evaluation and communication but also
can help improve models and increase flood resilience.
Keywords Crowdsourced observations, Data assimilation, Flood forecasting,
Hydraulic modelling, Hydrological modelling
1 Introduction
The impact of natural hazards on societies and economies has drastically increased in
the last years due to many natural and anthropogenic factors, including climate
change [1, 2]. For this reason, the demands for non-structural measures able to
accurately and timely forecast in real-time river water level to allow decision-makers
to take the most effective and timely decisions for reducing harm or loss have
significantly increased [3–5]. Among different types of water system models, hydrological and hydrodynamic models are the most utilised ones in flood early warning
systems in river basins.
Unfortunately, deterministic predictions contain an intrinsic uncertainty due to
many sources of error that propagate through the model and therefore affect its
output [6]. In fact, uncertainty can be due to either the inherent stochastic nature and
variability of hydrological processes, i.e. aleatory uncertainty [7, 8], or to our
imperfect state of knowledge of the hydrological system and our limitedness to
model it, i.e. epistemic uncertainty [9–12]. Three main sources of uncertainty can be
identified [13] in hydrological and hydrodynamic modelling: (a) observation uncertainty, which is the approximation in the observed hydrological variables used as
input or calibration data (e.g. rainfall, temperature and river discharge);
(b) parameter uncertainty, which is induced by imperfect model calibration; and
(c) model structural uncertainty, which is a result of the inability of models to
perfectly schematize the physical processes involved. Epistemic uncertainty can be
associated with the latest two sources of uncertainty previously mentioned due to
limited knowledge about the physical behaviour of the system.
A reliable characterisation and reduction of the uncertainties affecting hydrological and hydrodynamic processes is an important scientific and operational challenge
[14–17]. Different approaches like the first-order reliability method [18], probabilistic Monte Carlo (MC) and fuzzy rule-based methods [19–21] can be used to assess
model uncertainty.
210
M. Mazzoleni et al.
model structure, states and parameters. Two case studies, the Brue and Bacchiglione
catchments, are considered. Finally, the evaluation of the developed methods is
provided. This study demonstrates that networks of low-cost static and dynamic
social sensors can complement traditional networks of static physical sensors, for the
purpose of improving flood forecasting accuracy. This can be a potential application
of recent efforts to build citizen observatories of water, in which citizens not only can
play an active role in information capturing, evaluation and communication but also
can help improve models and increase flood resilience.
Keywords Crowdsourced observations, Data assimilation, Flood forecasting,
Hydraulic modelling, Hydrological modelling
1 Introduction
The impact of natural hazards on societies and economies has drastically increased in
the last years due to many natural and anthropogenic factors, including climate
change [1, 2]. For this reason, the demands for non-structural measures able to
accurately and timely forecast in real-time river water level to allow decision-makers
to take the most effective and timely decisions for reducing harm or loss have
significantly increased [3–5]. Among different types of water system models, hydrological and hydrodynamic models are the most utilised ones in flood early warning
systems in river basins.
Unfortunately, deterministic predictions contain an intrinsic uncertainty due to
many sources of error that propagate through the model and therefore affect its
output [6]. In fact, uncertainty can be due to either the inherent stochastic nature and
variability of hydrological processes, i.e. aleatory uncertainty [7, 8], or to our
imperfect state of knowledge of the hydrological system and our limitedness to
model it, i.e. epistemic uncertainty [9–12]. Three main sources of uncertainty can be
identified [13] in hydrological and hydrodynamic modelling: (a) observation uncertainty, which is the approximation in the observed hydrological variables used as
input or calibration data (e.g. rainfall, temperature and river discharge);
(b) parameter uncertainty, which is induced by imperfect model calibration; and
(c) model structural uncertainty, which is a result of the inability of models to
perfectly schematize the physical processes involved. Epistemic uncertainty can be
associated with the latest two sources of uncertainty previously mentioned due to
limited knowledge about the physical behaviour of the system.
A reliable characterisation and reduction of the uncertainties affecting hydrological and hydrodynamic processes is an important scientific and operational challenge
[14–17]. Different approaches like the first-order reliability method [18], probabilistic Monte Carlo (MC) and fuzzy rule-based methods [19–21] can be used to assess
model uncertainty.
210
M. Mazzoleni et al.