Several research activities aimed to reduce such uncertainty in the flood
estimation, predictive uncertainty, have been carried out due to its importance to
the decision of issuing a flood warning [5, 22, 23]. Methods like the UNcertainty
Estimation based on Local Errors and Clustering (UNEEC, [24–26]), Generalised
Likelihood Uncertainty Estimation (GLUE, [27, 28]) and Machine Learning in
parameter Uncertainty Estimation (MLUE, [29, 30]) can be employed to assess
uncertainty in water system models and estimate predictive uncertainty (see,
e.g. [31, 32]). However, such tools are often not used in operational forecasting by
environmental agencies and river basin authorities, perhaps because of their belief
that uncertainty analysis cannot be incorporated into the decision-making process
and because uncertainty analysis is too subjective, among others [5, 11, 33].
In the last decades, model updating techniques for reducing predictive uncertainty
approaches have been increasingly studied and implemented in water-related applications. These approaches allow for changing model input, states, parameters or
output in response of new observations coming into the model in order to improve
the prediction accuracy and quantifying uncertainty [3, 14, 34]. In most of the cases,
model updating occurs only in form of data assimilation using information of
streamflow, soil moisture, etc. coming from static physical stations. Model updating
techniques are rarely implemented in operational forecasting due to the lack of
approaches to quantify the uncertainty in real-time observations from multiple
sources across a range of spatiotemporal scales and methods to integrate these new
information in an appropriate and transparent way. In this respect, in operational
practice it is preferred to correct the model inputs (in most of the cases), states, initial
conditions and parameters in an empirical and subjective way rather than apply
advanced (optimal) data assimilation techniques for improving hydrologic forecast
[35]. Welles et al. [36] and Liu et al. [34] pointed out how the need for implementing
reliable data assimilation methods in operational forecast is increasing in order to fill
the mentioned gap with the scientific world.
Traditionally, static physical sensors, such as pressure sensors, water level sensors, and pluviometers, are commonly used by water authorities to calibrate, validate
and (in some cases) update physical models in real time. However, the main problem
of physical sensors is the proper maintenance which can be very expensive in case of
a vast network as well as the limited data that existing sparse monitoring networks
can provide to this end.
The continued technological advances have stimulated the spread of low-cost
sensors that has triggered crowdsourcing as a way to obtain observations of hydrological variables in a more distributed way than the classic static physical sensors
[37]. The main advantage of using these types of sensors is that they can be used not
only by technicians, as is the case of traditional physical sensors, but also by regular
citizens. Recently, citizen science activities have been widely promoted in order to
allow citizens to participate in different aspects of environmental planning and
management. One of the most common activities to achieve such goal includes
involving citizens in data collection, or crowdsourcing (CS). In particular, observations of hydrological variables can generate additional knowledge, in relation to the
water cycle, and use such knowledge in decision-making [38, 39]. However, because
Exploring Assimilation of Crowdsourcing Observations into Flood Models
211
estimation, predictive uncertainty, have been carried out due to its importance to
the decision of issuing a flood warning [5, 22, 23]. Methods like the UNcertainty
Estimation based on Local Errors and Clustering (UNEEC, [24–26]), Generalised
Likelihood Uncertainty Estimation (GLUE, [27, 28]) and Machine Learning in
parameter Uncertainty Estimation (MLUE, [29, 30]) can be employed to assess
uncertainty in water system models and estimate predictive uncertainty (see,
e.g. [31, 32]). However, such tools are often not used in operational forecasting by
environmental agencies and river basin authorities, perhaps because of their belief
that uncertainty analysis cannot be incorporated into the decision-making process
and because uncertainty analysis is too subjective, among others [5, 11, 33].
In the last decades, model updating techniques for reducing predictive uncertainty
approaches have been increasingly studied and implemented in water-related applications. These approaches allow for changing model input, states, parameters or
output in response of new observations coming into the model in order to improve
the prediction accuracy and quantifying uncertainty [3, 14, 34]. In most of the cases,
model updating occurs only in form of data assimilation using information of
streamflow, soil moisture, etc. coming from static physical stations. Model updating
techniques are rarely implemented in operational forecasting due to the lack of
approaches to quantify the uncertainty in real-time observations from multiple
sources across a range of spatiotemporal scales and methods to integrate these new
information in an appropriate and transparent way. In this respect, in operational
practice it is preferred to correct the model inputs (in most of the cases), states, initial
conditions and parameters in an empirical and subjective way rather than apply
advanced (optimal) data assimilation techniques for improving hydrologic forecast
[35]. Welles et al. [36] and Liu et al. [34] pointed out how the need for implementing
reliable data assimilation methods in operational forecast is increasing in order to fill
the mentioned gap with the scientific world.
Traditionally, static physical sensors, such as pressure sensors, water level sensors, and pluviometers, are commonly used by water authorities to calibrate, validate
and (in some cases) update physical models in real time. However, the main problem
of physical sensors is the proper maintenance which can be very expensive in case of
a vast network as well as the limited data that existing sparse monitoring networks
can provide to this end.
The continued technological advances have stimulated the spread of low-cost
sensors that has triggered crowdsourcing as a way to obtain observations of hydrological variables in a more distributed way than the classic static physical sensors
[37]. The main advantage of using these types of sensors is that they can be used not
only by technicians, as is the case of traditional physical sensors, but also by regular
citizens. Recently, citizen science activities have been widely promoted in order to
allow citizens to participate in different aspects of environmental planning and
management. One of the most common activities to achieve such goal includes
involving citizens in data collection, or crowdsourcing (CS). In particular, observations of hydrological variables can generate additional knowledge, in relation to the
water cycle, and use such knowledge in decision-making [38, 39]. However, because
Exploring Assimilation of Crowdsourcing Observations into Flood Models
211