Invited perspectives: How machine learning will change flood risk and impact assessment - Nat. Hazards Earth Syst. Sci.

Page created by Laurie Mejia
 
CONTINUE READING
Nat. Hazards Earth Syst. Sci., 20, 1149–1161, 2020
https://doi.org/10.5194/nhess-20-1149-2020
© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Invited perspectives: How machine learning will change flood
risk and impact assessment
Dennis Wagenaar1,2 , Alex Curran1,3 , Mariano Balbi4 , Alok Bhardwaj5 , Robert Soden6,7,8 , Emir Hartato9 ,
Gizem Mestav Sarica10 , Laddaporn Ruangpan11,3 , Giuseppe Molinario7 , and David Lallemant5,8
1 Department    of flood risk management, Deltares, Delft, the Netherlands
2 Institute for environmental studies, VU University, Amsterdam, the Netherlands
3 Faculty of Civil Engineering and Geosciences, Delft University of Technology, Delft, the Netherlands
4 Structural and Materials Lab, School of Engineering, Universidad de Buenos Aires, Buenos Aires, Argentina
5 Earth Observatory of Singapore, Nanyang Technological University, Singapore
6 Columbia University, New York City, New York, USA
7 GFDRR, World Bank Group, Washington, D.C., USA
8 Co-Risk Labs, Oakland, California, USA
9 Planet, San Francisco, USA
10 Institute of Catastrophe Risk Management, Nanyang Technological University, Singapore
11 Department of Water Resources and Ecosystems, IHE Delft Institute for Water Education, Delft, the Netherlands

Correspondence: Dennis Wagenaar (dennis.wagenaar@deltares.nl)

Received: 14 October 2019 – Discussion started: 16 October 2019
Revised: 17 February 2020 – Accepted: 15 March 2020 – Published: 29 April 2020

Abstract. Increasing amounts of data, together with more        machine learning is likely to drastically improve future flood
computing power and better machine learning algorithms to       risk and impact assessments, but issues such as applicability,
analyse the data, are causing changes in almost every aspect    bias and ethics must be considered carefully to avoid misuse.
of our lives. This trend is expected to continue as more data   This paper presents some of the current developments on the
keep becoming available, computing power keeps improv-          application of machine learning in this field and highlights
ing and machine learning algorithms keep improving as well.     some key needs and challenges.
Flood risk and impact assessments are also being influenced
by this trend, particularly in areas such as the development
of mitigation measures, emergency response preparation and
flood recovery planning. Machine learning methods have the      1   Introduction
potential to improve accuracy as well as reduce calculating
time and model development cost. It is expected that in the     Exponentially increasing computing power and data, as well
future more applications will become feasible and many pro-     as rapidly improving machine learning algorithms to anal-
cess models and traditional observation methods will be re-     yse these data, have been changing many aspects of our lives
placed by machine learning. Examples of this include the use    (Manyika et al., 2011). These trends are expected to con-
of machine learning on remote sensing data to estimate ex-      tinue and will undoubtedly keep affecting many scientific,
posure and on social media data to improve flood response.      commercial and social sectors (Manyika et al., 2011). Flood
Some improvements may require new data collection efforts,      risk and impact assessments are no exception to this trend.
such as for the modelling of flood damages or defence fail-     Flooding yearly affects more people than any other natural
ures. In other components, machine learning may not always      hazard type (Jonkman, 2005), and the impact and frequency
be suitable or should be applied complementary to process       of flooding events is expected to increase in the future due
models, for example in hydrodynamic applications. Overall,      to urban development and climate change (Kundzewicz et
                                                                al., 2014). It is therefore an opportunity for researchers and

Published by Copernicus Publications on behalf of the European Geosciences Union.
1150       D. Wagenaar et al.: Invited perspectives: How machine learning will change flood risk and impact assessment

Table 1. Overview of different types of flood risk and impact as-
sessments.

              Predictive             Descriptive
  Exposure    Urban growth mod-      Identification of current
              elling                 built-up area
  Hazard      Flood modelling        Mapping current and past
                                     floods
  Impact      Forecasting impact     Assessing flood impacts
              (e.g. damage)          (e.g. damage) after they
                                     have occurred
                                                                    Figure 1. Disaster management cycle, a common paradigm tool.

flood managers to tap into the potential of machine learning,
taking advantage of their strengths while being cognizant of        used (e.g. Wagenaar et al., 2019), typically for the design of
their limitations. It is also important to anticipate improve-      risk-reduction interventions ranging from protective infras-
ments in the capabilities of machine learning methods, so as        tructure to insurance products. The challenge in this phase
to plan for forthcoming changes in flood modelling.                 is model reliability and uncertainties about future develop-
   When assessing the interaction between floods and soci-          ments (e.g. uncertainty in future exposure). In the preparation
ety, three different components can be recognized: exposure,        phase, predictive models are used for emergency planning
hazard and impact (Kron, 2002). Exposure refers to the char-        (e.g. Coughlan de Perez et al., 2016), where the challenge is
acteristics of the people and assets that can be affected by        the reliability, availability and communication of data. Ma-
flooding. Hazards are the physical characteristics of a flood       chine learning is capable of generating more reliable and
such as the extent, water depth, duration and flow velocity.        faster models that can help solve some of the current chal-
Impacts are the effects the hazard has on the exposure. To          lenges in the disaster management cycle but could also pro-
assess these three components, we make the distinction be-          vide new opportunities (GFDRR, 2018).
tween flood risk, as the probabilistic analysis of the potential       Machine learning algorithms can find patterns in data
(predictive) impacts of floods, and flood impact assessment,        and use these patterns to make predictions about new data
as the post-event assessment of (descriptive) impact from an        (Bishop, 2006). For example, when a machine learning algo-
actual flood event. Table 1 provides examples of predictive         rithm is provided with aerial images of either urban or rural
and descriptive assessments in relation to the hazard, expo-        areas and corresponding labels (urban or rural), it can build
sure and impact components. The scope of this paper is lim-         the capacity to classify new unlabelled aerial images as ei-
ited to the predictive and descriptive assessments shown in         ther urban or rural. Features in the above example would be
Table 1 and does not include potential uses of machine learn-       different components of the aerial images (i.e. pixel tone and
ing in risk awareness or communication strategies.                  locations), and the target variable would be the label (i.e. ur-
   Flood risk and impact assessments have many differ-              ban or rural). When a precise value is required as opposed to
ent applications. A useful paradigm through which look at           a label, it is called a “regression task” (e.g. Bishop, 2006).
these different applications is the “disaster management cy-        An example of this is in flood damage modelling, where fea-
cle” (Khan et al., 2008; National Research Council, 2006)           tures such as water depth, flow velocity and building materi-
(Fig. 1). This cycle delineates the phases between events, i.e.     als can be used to predict a target variable such as monetary
the immediate response to an event, the long-term recovery,         economic damage based on historical records (e.g. Merz et
the mitigation to prevent future events and the preparation         al., 2013; Wagenaar et al., 2017). Due to the use of labelled
prior to a new forecasted event.                                    training data (e.g. classified images or historic damage ex-
   In the response phase, the focus is typically on descrip-        amples), regression and classification are called supervised
tive hazard, exposure and impact assessments (e.g. Klemas,          learning tasks. Machine learning method categories also in-
2015), sometimes complemented with predictive models if             clude unsupervised learning and reinforcement learning (see
the event descriptive information is not available yet (e.g. a      GFDRR, 2018). However, such methods are not discussed in
predictive model estimating the number of people affected           this paper because they are expected to have a smaller short-
can be fed by a descriptive hazard model of the flood ex-           term impact on the field of flood risk and impact assessments.
tent). The challenge in this phase is mostly data reliability.         The simplest machine learning algorithms have been used
In the recovery phase, descriptive assessments are often used       for a long time and are often known as basic statistical
for payouts (e.g. indemnity insurance), and one of the main         techniques (e.g. linear regression: Legendre, 1805; Gauss,
challenges is ensuring these payouts are timely and reliable.       1809). More sophisticated machine learning techniques that
In the mitigation phase, probabilistic predictive models are        emerged in the 1980s and 1990s (e.g. decision trees and

Nat. Hazards Earth Syst. Sci., 20, 1149–1161, 2020                          www.nat-hazards-earth-syst-sci.net/20/1149/2020/
D. Wagenaar et al.: Invited perspectives: How machine learning will change flood risk and impact assessment                    1151

neural networks) can find more complex non-linear patterns           2     Perspective per component
(Breimann et al., 1984; Rumelhart et al., 1986). Recent ad-
vances in machine learning (e.g. convolutional neural net-           2.1     Exposure assessment
works) make computer vision and other advanced applica-
tions possible (Krizhevsky et al., 2012). The more advanced          2.1.1    Descriptive exposure assessments
techniques such as decision trees, neural networks and espe-
cially convolutional neural networks can find more complex           Descriptive exposure assessments consist of detecting and
patterns. This is because they allow for more complex non-           characterizing (spatial) features such as current buildings,
linear functions to be fitted to the data. Such complex func-        agriculture fields, roads and other infrastructure. Tradition-
tions require a large number of model coefficients to be set         ally this has been done by population censuses, building
during the training of the model. To set all these coefficients,     counts and conventional mapping techniques that require
a lot of training examples are required. In some cases the           ground surveys. Remote sensing is currently changing this.
number of training examples can be reduced with transfer             It has become common for aerial and satellite images to be
learning techniques (Olivas et al., 2010). These techniques          manually digitized and labelled to make building footprints
make it possible to re-use knowledge gained from other prob-         or map roads. This has been done by “crowds” of mappers
lems to train a model on a smaller training data set.                in “mapathons”, for example using the OpenStreetMap plat-
   From the beginning, machine learning has been used in             form. Machine learning is very likely going to drastically
predictive flood hazard modelling (Solomatine and Ostfield,          change this. Research into automatically labelling remote
2008) mostly as a faster and simpler alternative to process          sensing data has already been going on for some time (e.g.
models. A simple example of this is the prediction of river          Heermann and Khazenie, 1992; Giacinto and Roli, 2001). It
discharge based on upstream rainfall data (e.g. Dibike and           is already being used to label build-up areas based on night-
Solomatine, 2001). This type of modelling has been practised         time lights (Goldblatt et al., 2018) or satellite images (Gold-
for a long time but has not displaced the traditional process        blatt et al., 2016). Furthermore, algorithms are already being
models. This is probably because the methods are not suffi-          used to automatically label buildings (Sermanet et al., 2014;
ciently better than traditional methods to offset some disad-        Alshehhi et al., 2017; GFDRR, 2018) and map roads (Gao et
vantages as discussed in the predictive-hazard section. In re-       al., 2019) using aerial/satellite imagery. This will reduce the
cent years, more data have become available through remote           need for manual detection and will probably provide global
sensing, social media (e.g. Fohringer et al., 2015), citizen sci-    availability of such building footprints and road information
ence (e.g. Annis and Nardi, 2019) and other sources. This im-        in the near future.
pulse of new data combined with machine algorithms could                Part of an exposure assessment is the observation of as-
lead to changes in flood risk and impact assessment. Some            set features relevant to risk analysis, for example building
of these changes have already been highlighted by major in-          materials, building occupancy (e.g. residential or industrial),
ternational organizations such as the World Bank (GFDRR,             building height, ground floor elevation, poverty rates in the
2018).                                                               population etc. This information is typically not available but
   This invited perspective paper starts with a perspective per      could be very valuable as input for impact models (e.g. Merz
risk assessment component as defined in Table 1. These spe-          et al., 2013; Wagenaar et al., 2017; Schröter et al., 2014) or,
cific perspectives start with a description of the traditional ap-   for example, to account for poverty in cost–benefit analyses
proach for the assessments, followed by a literature review on       (e.g. Kind et al., 2016). Similarly, ground floor elevation in-
how machine learning techniques are currently being devel-           formation could radically improve urban pluvial flood dam-
oped to improve the traditional approach, and then proceed to        age modelling as damage from small-scale floods is very sen-
speculate on potential future improvements. This is followed         sitive to such variables.
by a “General perspectives” chapter, in which general trends            Some work has already been carried out on detecting
that come back in the different components are identified and        poverty (Watmough et al., 2019) and building heights (Saadi
discussed. This includes common challenges (i.e. data limi-          and Bensaibi, 2014) by satellite imagery. Another source of
tations, transferability, ethics and bias) and ends with some        this building feature information could be 360◦ street view
speculation about the likelihood of future developments.             images combined with computer vision techniques. Such im-
                                                                     ages are available in, for example, the open-source street
                                                                     view data platform Mapillary (Neuhold et al., 2017). Such
                                                                     techniques are already starting to impact earthquake risk as-
                                                                     sessments, such as in Guatemala, where 360◦ imagery was
                                                                     fed into Mapillary algorithms in order to automatically de-
                                                                     tect “soft story” buildings: those most likely to collapse in an
                                                                     earthquake. This was done by having the machine learning
                                                                     algorithm detect features that were indicators of large open-
                                                                     ings on the ground floor of buildings (large doors, garage

www.nat-hazards-earth-syst-sci.net/20/1149/2020/                             Nat. Hazards Earth Syst. Sci., 20, 1149–1161, 2020
1152      D. Wagenaar et al.: Invited perspectives: How machine learning will change flood risk and impact assessment

doors, shop windows etc.) (GFDRR, 2018). Computer vision             future land-use maps and make high-resolution future land-
techniques from street level imagery are currently limited to        use maps globally available.
detecting such relatively simple features. However, based on
recent advances seen in other computer vision applications           2.2     Hazards assessment
(e.g. facial recognition), it is likely that in the future it will
be possible to detect more complex building features as well.        2.2.1    Descriptive hazard assessments
For computer vision models to detect complex information
like ground floor elevation or building materials, it would          Descriptive flood hazard assessment focuses primarily on the
be necessary to provide labelled examples to the algorithms.         response phase, i.e. estimating current inundation extents and
Such labelled examples are in some cases already available           depths to assist both emergency responders and those af-
for some areas, e.g. ground floor elevation (Bouwer et al.,          fected directly. This is traditionally achieved using optical
2017) or building materials (Schröter et al., 2018).                 remote sensing data, local sensors or manually collected data
                                                                     from observers on the ground. However, the rise of two major
                                                                     data sources, synthetic aperture radar (SAR) and social me-
2.1.2   Predictive exposure assessments
                                                                     dia, provides a number of opportunities for machine learning
                                                                     to improve upon current flood detection methods.
Predictive exposure mapping consists of estimates of future             During a flood event, affected populations frequently pro-
exposure. This mostly includes modelling to predict urban            duce “user-generated content” or “crowd-sourced” data from
growth and other changes in land use. It is required for eval-       social media posts or apps where citizens can report floods
uating flood mitigation measures (e.g. Wagenaar et al., 2019)        (Mazoleni et al., 2017; Assumpção et al., 2018; Annis and
because such measures typically need to function for a long          Nardi, 2019; UrbanRiskLab, 2019). This is especially the
time and should therefore still perform as required after pre-       case in urban areas, where internet and social media penetra-
dicted land-use changes. Land-use changes affect the impact          tion are higher compared to rural areas. These data are often
of a flood because more damage may occur for the same                “tagged” temporally and spatially and can be used by ma-
flood hazard and the flood hazard may become greater be-             chine learning algorithms for applications such as nowcast-
cause of changes in impervious area and therefore rainfall           ing by searching for certain keywords like “flood” (e.g. see
runoff (Triantakonstantis and Mountrakis, 2013; Mestav Sar-          Tkachenko et al., 2017; Bischke et al., 2017; Lopez-Fuentes
ica et al., 2019). Predictive exposure assessments for flood         et al., 2017). The method is currently used to map real-time
risk and impact assessments are currently often not carried          flood extents in several countries (Eilander et al., 2016). Po-
out spatially, but rather GDP growth projections are applied         tential future machine learning and computer vision tech-
to estimate future total exposure values (e.g. van der Most et       niques could be extended to estimate water depths and other
al., 2014; Wagenaar et al., 2019). This is enough for some           flood characteristics from posted photos.
studies, but if large changes are expected a land-use change            Remotely sensed optical data are often used to identify
or urban growth model is required.                                   the extents of flooding, but optical sensors are not functional
   Urban growth has been modelled with simple ma-                    during periods of cloud cover or at night. Furthermore, the
chine learning models in the past (e.g. logistic regression)         temporal resolution often prevents the observation of flash
(Samardzic-Petrovic et al., 2017). The use of cellular au-           floods. SAR data using the microwave wavelengths of the
tomata (CA) models has become more common recently                   electromagnetic spectrum can help overcome these problems
(Naghibi et al., 2016). These models assign cells as either          by providing additional imagery during the night or during
urban or non-urban based on specific transition rules. De-           cloud cover. Adding night-time and cloud-cover images will
termining the optimum transition rules is a critical issue for       provide a higher total temporal resolution. Flood extents are
CA modelling (Aarthi and Gnanappazham, 2019). This is                currently determined with statistical methods using thresh-
sometimes difficult because of human bias, heterogeneity             olds to subsequently identify flood extents, e.g. by using
and non-linear relations between driving factors and urban           Bayesian method on SAR amplitude time-series data (Lin et
expansion (Naghibi et al., 2016; Xu et al., 2019). To over-          al., 2019). Advanced machine learning classification meth-
come these limitations, machine learning algorithms such             ods are being developed to improve this process, but in order
as artificial neural networks have been integrated with tra-         to train them it is necessary to have manually labelled im-
ditional CA to model urban growth (Aarthi and Gnanap-                ages as training data. Collection of this labelled flood extent
pazham, 2019; Naghibi et al., 2016). They then use historical        information is the main challenge for automatic detection
land-use changes (e.g. Song et al., 2015) to learn the transi-       moving forward. Manual methods could harness the power
tion rules. Complex machine learning models have also been           of the crowd, as people are connected through the inter-
directly applied to urban growth modelling without the CA            net or with mapathons. These approaches could have game-
model structure (Pal and Ghosh, 2017). These improvements,           changing implications for the training of machine learning al-
together with more data about past land-use changes and ad-          gorithms. Already mapathons are often “trainathons”, where
ditional computation power, are expected to provide better           mappers are not only manual digitizers but also labellers and

Nat. Hazards Earth Syst. Sci., 20, 1149–1161, 2020                            www.nat-hazards-earth-syst-sci.net/20/1149/2020/
D. Wagenaar et al.: Invited perspectives: How machine learning will change flood risk and impact assessment                 1153

trainers for automated machine learning methods for the fu-       machine learning methods to optimize control decisions in
ture.                                                             the event of communication network breakdowns during ex-
                                                                  treme storm events.
2.2.2   Predictive hazard assessments                                Another major application for machine learning in long-
                                                                  term risk analysis is “surrogate” modelling (Ong et al.,
Predictive flood hazard assessments consist of predicting fu-     2003), in which the outputs from process models are used to
ture floods and their characteristics such as extents, inunda-    train computationally less-intensive machine learning mod-
tion depths, durations, flow velocities, waves and water lev-     els. This can be applied to speed up different types of pro-
els in rivers or seas. These assessments are applied for short-   cess models applied in predictive hazard modelling. For ex-
term forecasting in the preparation phase (preparing for im-      ample, in flood defence analysis and design, classical reli-
minent events) and long-term risk analyses for use in flood       ability techniques such as the first-order reliability method
risk management (mitigation phase).                               (FORM) and Monte Carlo simulations (Steenbergen et al.,
   In flood forecasting, traditional methods of predicting haz-   2004), or large-scale risk analyses that utilize them (Curran et
ard variables can involve a chain of hydrologic and hydraulic     al., 2019), can be replicated using a relatively small amount
models that describe the physical processes. Although such        of evaluations as samples (Chojaczyk et al., 2015; Kingston
models provide system understanding, they often have high         et al., 2011). However, surrogate models may be particularly
computational and data requirements. Therefore, the use of        susceptible to extrapolation problems, where input data out-
process models may not always be feasible or necessary in         side the range of the training data are introduced (Ghalkhani
the preparation stage of a disaster. At that moment, accu-        et al., 2013).
rate and timely outputs become more important than system            In the mitigation phase a chain of hydrologic and hydraulic
understanding, and the use of “black-box” machine learn-          models that describe the physical processes is typically ap-
ing models (e.g. Campolo et al., 2003) is becoming more           plied (e.g. Wagenaar et al., 2019). In general, system under-
widespread (Mosavi et al., 2018). The increased speed can         standing is required to assess proposed or potential future
create a trade-off with the robustness of forecast models, as     changes. In such cases, data-driven approaches are typically
changes to the hydraulic system (such as a new structure that     not applicable as there are no data about how the system be-
could be easily implemented into a hydraulic model) cannot        haves after the changes occur, and hence simulation models
be directly introduced into a trained machine learning model.     are required that describe the physical system.
In addition, machine learning models might not perform well
in predicting extremes far outside past observations, since       2.3     Flood impact assessment
they have not been trained against such extremes.
   A review of flood forecasting methods using machine            2.3.1    Descriptive impact assessments
learning by Mosavi et al. (2018) highlights trends such as
component and ensemble models (collectively termed “hy-           Descriptive impact assessments consist of making estimates
brid models”; Corzo and Solomatine, 2014). Hybrid compo-          of the flood impact after or during an event. This is tradition-
nent models assign machine learning a specific task in the        ally done with manually collected data from observers on the
modelling process that is either highly complex or not well       ground. However, such manual ground inspections are slow
understood. Examples of this include using machine learning       and require people to enter the disaster area. Remote sens-
for error correctors (see, for example, studies by Abrahart       ing can be used to get a very quick first impression of the
and See, 2007, and Google Research – Nevo et al., 2019)           damage to help with disaster response. Such techniques have
or flows subject to human influence (Yaseen et al., 2019).        already been applied, for earthquake and wind damage (e.g.
Hybrid ensemble methods often use machine learning mod-           Menderes et al., 2015). For flooding, this is often more dif-
els to supplement process models, providing robust predic-        ficult because damage inside buildings is difficult to obtain
tions and uncertainty ranges (Solomatine and Ostfeld, 2008).      either from aerial or space-based sensors. Only when build-
Such methods benefit from the speed and ability to deal with      ings completely collapse or are removed by strong flows does
non-linear multivariable problems of machine learning mod-        remote sensing become feasible. This is, for example, the
elling and the process understanding available in conven-         case with flash floods, tsunamis or some storm surges. If 360◦
tional modelling. The review by Mosavi et al. (2018) does         street view images are collected after a flood, these could po-
not consider gridded/spatial forecasting techniques, but ad-      tentially be used for damage assessment. Machine learning
vanced machine learning techniques are starting to be de-         techniques could then eventually be used to give a quick first
veloped for precipitation pattern nowcasting (Xingjian et         estimate of the damage.
al., 2015) and flood extents prediction (Chang et al., 2018).        The use of machine learning techniques for automatic de-
Another application of machine learning in the preparation        tection of damages from remote sensing information (aerial
phase is in the real-time control of flood defences and sys-      or street view) requires labelled training data from manu-
tems (e.g. Lobbrecht and Solomatine, 2002; Castelletti et al.,    ally collected data from observers on the ground. These data
2010). For example, Lobbrecht and Solomatine (2002) used          are currently rare. An approach could be to start using re-

www.nat-hazards-earth-syst-sci.net/20/1149/2020/                           Nat. Hazards Earth Syst. Sci., 20, 1149–1161, 2020
1154      D. Wagenaar et al.: Invited perspectives: How machine learning will change flood risk and impact assessment

mote sensing data to manually label the impact. A way to get          Machine learning could also be applied to predict disease
around this limitation is to detect changes pre- and post-flood    outbreak after floods by combining remote sensing, meteoro-
using high-resolution satellite images for urban areas where       logical and socio-economic data (e.g. Mayfield et al., 2018;
many buildings are damaged. Pixels with changed informa-           Carvajal et al., 2018; Modu et al., 2017; Yomwan et al., 2015;
tion will denote the damage that happened due to the floods.       Tiwari et al., 2013; Shively et al., 2015). In a flood event,
Eventually these data can then be used as training data for        there is an increased risk of infectious diseases among sur-
cases where only the post-flood images are available within a      vivors and displaced persons. For example, measles, diar-
short time interval after the flood event. This method would       rhea, acute respiratory infections and malaria can be respon-
however only be relevant to catastrophic floods because it         sible for many deaths (Lignon, 2006). Predictive modelling
does not address the fact that most damage remains not ob-         of such diseases is rarely carried out, and current approaches
servable from top view. On top of that, this approach intro-       mostly focus on simple regression models or process models
duces significant new error: (1) error in the change detection     that simulate the spread of pollutants in the water. One major
signal, (2) error in relating the change to damage and (3) error   challenge is that the degree to which such epidemics occur is
in training a new model based on those damage labels. Im-          associated with the regional endemicity of specific diseases,
agery from different angles (e.g. from street view or drones)      the nature and scope of the disaster, the level of public health
might be more useful for change detection; however these           infrastructure in place both before and after the event, and
data would also be more difficult to acquire.                      the level and efficacy of disaster response (Ivers and Ryan,
                                                                   2006). Machine learning models could take such complex
2.3.2   Predictive impact assessments                              processes better into account.
                                                                      Machine learning can be used for structural health moni-
Predictive flood impact assessments include models that            toring; this has applications in the preparation phase (Pyayt
translate hazard and exposure information into socio-              et al., 2014; Jonkman et al., 2018) and in the long-term reli-
economic impacts of the flood. This can include informa-           ability analysis required in the mitigation phase (Prendergast
tion such as monetary flood damage, casualties, buildings          et al., 2018; Klerk et al., 2019). In the preparation phase for
damaged, crop damage, disease outbreak, building materials         a flood, structural health monitoring is often done by manual
needed, recovery time, health monitoring of key structures         inspections of the infrastructure on the ground. For example,
and indirect damage (damages that occur in a different spa-        in the Netherlands there is a large network of volunteers that
tial and/or temporal setting than the originating event).          can be activated in the event of high river levels to inspect
   Most predictive flood damage modelling relies on depth–         the dikes. In the mitigation phase this is done by geotech-
damage functions that describe a relationship between the          nical process models fed by observations from the ground
water depth and monetary flood damage (Merz et al., 2010).         (e.g. De Waal, 2016); this is for example applied to decide
They are based either on historical flood damage records           on dike strengthening. Machine learning algorithms can help
(e.g. Thieken et al., 2008; Kreibich et al., 2010) or on ex-       detect damage patterns from sensor data and are currently be-
pert estimates (e.g. Penning-Rowsell et al., 2005). In prac-       ing used for the monitoring of flood defence structures such
tice, many more variables than water depth have an influence       as dikes (Pyayt et al., 2011). Similar methods have also been
on the flood damage (Cammerer et al., 2013; Wagenaar et al.,       applied to bridges (Neves et al., 2017). The use of both ma-
2016). Therefore, in the scientific literature there has been      chine learning algorithms and traditional techniques for dam-
a move towards multivariable flood damage models that use          age detection during floods is still very scarce (Prendergast et
many variables (e.g. flood duration, velocity, building materi-    al., 2018; Pyayt et al., 2011); however, integration of struc-
als, socio-economic status of inhabitants etc.) instead of just    tural health monitoring with flood early-warning systems is
water depth (e.g. Merz et al., 2013; Spekkers et al., 2014;        a very promising field of development for machine learning
Chinh et al., 2015; Kreibich et al., 2017; Wagenaar et al.,        techniques but would also require training data.
2017; Carisi et al., 2018; Amadio et al., 2019). These mod-           Indirect damages and business interruption are often taken
els are based on data and machine learning. The problem lies       into account simply through a scaling factor of the direct
with insufficient data availability to train machine learning      damage (e.g. Wagenaar et al., 2019). More complex mod-
models and with the fact that using the models requires a          els for quantifying such damages include input–output mod-
lot of feature data about flood and building characteristics       els and general equilibrium models (e.g. Koks et al., 2016).
plus socio-economic data about inhabitants (Wagenaar et al.,       To quantify indirect damages, such as business interruption
2017). In the future we expect more data about features to be-     losses, estimating the time it will take for different assets to
come available from computer vision applied to street view,        be back in full or partial functionality is required. These post-
satellite or drone images (see descriptive exposure section).      disaster restoration models have started to be formalized in
This would improve the quality of such models, could make          the last few years, primarily focused on earthquake disasters
it easier to apply them and make the development possible          (Kang et al., 2018; Burton et al., 2018). Due to a lack of gath-
for more areas.                                                    ered empirical data on post-disaster recovery, the use of data-
                                                                   intensive machine learning techniques has not yet made an

Nat. Hazards Earth Syst. Sci., 20, 1149–1161, 2020                         www.nat-hazards-earth-syst-sci.net/20/1149/2020/
D. Wagenaar et al.: Invited perspectives: How machine learning will change flood risk and impact assessment                  1155

impact on this discipline. However, the need to probabilisti-      A number of technology companies and research institutions
cally quantify recovery will require the use of statistical mod-   have developed guidelines for evaluating machine learning
els for calibration or assessments of recovery times, and that     systems, but this work is still evolving. Despite similar poten-
might be possible in the near future with the use of new re-       tial for negative impacts of these tools in flood risk manage-
mote sensing and crowd-sourcing technologies to obtain the         ment (Soden et al., 2019), the community has not given these
empirical feature data needed.                                     issues as much attention. Such concerns include the potential
                                                                   for reinforcing existing social inequalities and the reduced
                                                                   role of human judgement in modelling processes. These are
3     General perspectives                                         risks that need to be weighed seriously against the potential
                                                                   benefits of machine learning and explored in greater detail
3.1    Data limitations                                               Biases in machine learning can occur because of data sets
                                                                   that, for a number of reasons, do not fully represent the phe-
Many machine learning applications in flood risk and impact
                                                                   nomena which they are meant to describe (e.g. people are
modelling appear to be limited by a lack of data, especially
                                                                   accidentally excluded). For example, we often measure what
training data needed to build effective machine learning mod-
                                                                   we have data for, rather than measuring what matters most, or
els. This is especially true since the field of flood risk anal-
                                                                   use training data sets that reinforce past problems. For exam-
ysis is concerned primarily with extreme events, which are
                                                                   ple, if certain settlements are not detected in exposure maps,
rare, and data-collection during such events is often logisti-
                                                                   because they use different construction practices than the
cally difficult. The increase in the amount of data around the
                                                                   settlements used in training data sets, they may not receive
world does not necessarily imply that this problem will be
                                                                   emergency aid in the event of a flood. These problems can
resolved in the future. Some data are simply not collected, or
                                                                   be mitigated by ensuring modellers understand the context
there are measurement definition or quality issues. To fulfil
                                                                   of what they are attempting to model. Other ethical issues
the potential of machine learning, new data collection efforts
                                                                   raised by machine learning in the flood management con-
will be required, along with data standardization protocols.
                                                                   text include data ownership, transparency, consent and pri-
This will necessitate collaboration between different organi-
                                                                   vacy. For example, some people may object to having their
zations and stakeholders, setting of data standards and a will-
                                                                   home labelled “vulnerable” on a vulnerability map used by
ingness to share. This problem is common to impact data col-
                                                                   first responders. Privacy concerns may be aggravated by ma-
lected (see Sect. 2.3.1 and 2.3.2), labelled flood extent data
                                                                   chine learning and other big-data techniques. Ethics prob-
(see Sect. 2.2.1), social media hazard data (see Sect. 2.2.1)
                                                                   lems should be addressed by carefully weighing the benefits
and first-floor elevation data (see Sect. 2.1.1).
                                                                   of collecting certain data against the related privacy costs, in
3.2    Transferability of data                                     collaboration with people who may be affected by the out-
                                                                   comes of decisions based on machine learning tools.
A critical assumption behind machine learning techniques              An additional ethical concern regarding machine learn-
is that the data being used to train a model are representa-       ing in flood risk assessment is misuse of models. In some
tive of the situation the model needs to be applied in. For        sectors great advances have been made with machine learn-
example, a data set on damage to concrete buildings is not         ing (e.g. facial recognition and self-driving cars). This suc-
fully applicable to modelling the damage to thatched huts. It      cess for some tasks can lead to an awe-inspiring general atti-
is therefore important to collect heterogenous data sets that      tude towards the techniques (Ames, 2018; Narayanan, 2019).
cover a large spectrum of potential situations (Wagenaar et        This hype sometimes leads to unwarranted trust in the tech-
al., 2018). Data that are not fully applicable can still have      niques for tasks machine learning is not (yet) suitable for
some value, for example through domain adaptation or trans-        (Narayanan, 2019). For example, many companies are cur-
fer learning (GFDRR, 2018), but applicable data are always         rently using machine learning for hiring decisions despite
required as well. Wagenaar et al. (2020) showed that sample        well-documented failings of these tools (Narayanan, 2019;
selection bias correction, a form of domain adaptation, helps      Raghavan et al., 2019). In order to avoid such misuse in flood
to improve machine learning impact models in a transfer set-       risk assessment, it is important that machine learning im-
ting. Furthermore, it is important to work on efficient ways       plementations are transparent and supervised by independent
to communicate the applicability of data-driven models.            machine learning and flood risk assessment experts.
                                                                      Importantly, flood risk assessments are highly data re-
3.3    Ethics and bias                                             liant, and the increased attention to questions of ethics and
                                                                   bias in machine learning systems might serve as an op-
Significant attention is currently being given to questions of     portunity to drive conversations in our field about the lim-
the ethics and bias of machine learning systems across a va-       its of disaster data more broadly. Many of the sources of
riety of domains, including facial recognition (Keyes, 2018),      bias or ethical concerns in machine learning systems orig-
automated weaponry (Suchman and Weber, 2016), criminal             inate in, or share common roots with, other kinds of data
justice (Eubanks, 2018) and search engines (Noble, 2018).          used to understand disaster risks. This includes issues such

www.nat-hazards-earth-syst-sci.net/20/1149/2020/                          Nat. Hazards Earth Syst. Sci., 20, 1149–1161, 2020
1156       D. Wagenaar et al.: Invited perspectives: How machine learning will change flood risk and impact assessment

Table 2. Future predictions.

                               Predictive                                       Descriptive
                  Exposure     Likely incremental changes, e.g. improved        Very likely significant changes, e.g. auto-
                               cellular automata transition rules               matic exposure detection including building
                                                                                features
                  Hazard       Diverse field; changes are more likely to be     Likely changes in detection due to remote
                               complementary or specific components             sensing and social media algorithms.
                               of modelling
                  Impact       Potential for significant changes (i.e. multi-   Significant changes likely for some ele-
                               variable data-driven methods)                    ments; others will probably remain the same

as (1) property values driving what areas get protected,                   nomic feasibility, which is difficult to assess, combined with
(2) privacy concerns (which may be aggravated by machine                   conservative users. An example of this is the large-scale col-
learning and other big-data techniques), (3) how the lack                  lection of impact data which is required for both descriptive
of gender/age/ethnicity-disaggregated data on disaster risk                and predictive impact modelling (see Sect. 2.3.1 and 2.3.3)
masks differential vulnerabilities and (4) the importance of               or the training data required for descriptive hazard assess-
public participation and the voice of residents of areas por-              ments (see Sect. 2.2.1). Sometimes the obstacle is also tech-
trayed by models as “at risk”. Detailed analyses of specific               nical feasibility, for example whether it will really be possi-
cases (e.g. Soden and Kauffman, 2019) are urgently needed                  ble to extract first-floor elevation levels from street view (see
to make further progress in understanding the consequences                 Sect. 2.1.2). Innovations are also interdependent; for exam-
of the assessment methods we use to understand disasters.                  ple, when building feature information can be automatically
                                                                           extracted from street view, impact models will become easier
3.4     Future predictions                                                 to train and easier to run, and it will make more sense to start
                                                                           collecting the required impact data.
In the following section we draw some general conclusions
about how machine learning will change flood risk and im-                  3.4.3   Unlikely changes
pact assessments. Table 2 provides an overview of these pre-
dictions.                                                                  For some processes, machine learning may not be the best so-
                                                                           lution from a theoretical perspective. For example, the pro-
3.4.1    Very likely changes
                                                                           cesses of how water flows are very well known and can be
A few of the trends seem inevitable, primarily in cases where              well approximated with existing equations. It, therefore, does
recent technological advances or data that recently became                 not always make sense to pick a machine learning approach.
available make next steps obvious. A good example of this is               Another situation in which machine learning is not applica-
the automatic detection of building footprints and roads from              ble is when a system is being modelled on which predictions
high-resolution remote sensing imagery (see Sect. 2.1.1).                  need to be made that cannot have been seen in the data or
This is already possible and will, especially in data-poor                 when we know from an exploratory data analysis that we
areas, drastically improve the quality of the first response               have no data for it (GFDRR, 2018), for example how a sys-
and risk calculation. Further advances in the use of machine               tem may behave under never-seen discharges or after new
learning in descriptive hazard assessment through social me-               infrastructure has been built (e.g. new dam in the river). In
dia are also inevitable (see Sect. 2.2.1), given the amount of             these cases, machine learning may play a role in some com-
data available to social media companies.                                  ponents of the model, but process models will very likely re-
                                                                           main crucial in simulating the never-before-seen conditions.
3.4.2    Likely and potential changes                                      Especially for predictive hazard models (see Sect. 2.2.2),
                                                                           there are many elements that are unlikely to change with the
This is the category that can be shaped the most by individual             advance of machine learning.
innovators, and the majority of the advances discussed in this
paper fall under this category. In this case, the innovation still         3.4.4   New practices in flood risk and impact
experiences some kind of obstacle that prevents widespread                         assessments
application. It is typically difficult to predict whether such
obstacles can be truly removed in the future and how long                  Most change to flood risk and impact assessments discussed
that will take. Because the field of flood risk and impact                 in this paper relate to better models. Such cheaper, faster and
assessments is relatively small, the obstacles are often eco-              more accurate models could possibly yield new practices in

Nat. Hazards Earth Syst. Sci., 20, 1149–1161, 2020                                 www.nat-hazards-earth-syst-sci.net/20/1149/2020/
D. Wagenaar et al.: Invited perspectives: How machine learning will change flood risk and impact assessment                          1157

flood risk and impact assessments. Cheaper models would                Financial support. This research has been supported in part by
make flood risk and impact assessments feasible to carry out           Deltares and the National Research Foundation, Prime Minister’s
for a larger group of users and are therefore likely to make           Office, Singapore, under the NRF-NRFF2018-06 award.
emergency aid and investments in mitigation measures more
efficient. Faster methods may speed up emergency response
and recovery, especially when manually collected data from             Review statement. This paper was edited by Heidi Kreibich and re-
observers on the ground are replaced by remote earth obser-            viewed by Fernando Nardi and one anonymous referee.
vation. More accurate models may lead to more early actions
being feasible (Coughlan de Perez et al., 2016), and hence
early actions can be carried out that could not be carried             References
out before, for example more targeted measures during the
preparation and response phase of a flood. Such new mea-               Aarthi, A. D. and Gnanappazham, L.: Comparison of Urban Growth
sures include providing emergency payouts even before the                Modeling Using Deep Belief and Neural Network Based Cellular
event to the most vulnerable people (e.g. Reuters, 2019), pri-           Automata Model – A Case Study of Chennai Metropolitan Area,
oritization of emergency measures in buildings, targeted dis-            Tamil Nadu, India, Journal of Geographic Information System,
                                                                         11, 1–16, 2019.
ease outbreak prevention (Coughlan de Perez et al., 2016),
                                                                       Abrahart, R. J. and See, L. M.: Neural network modelling of non-
early shipping of the right emergency goods (Coughlan de                 linear hydrological relationships, Hydrol. Earth Syst. Sci., 11,
Perez et al., 2016) and prioritization of early harvesting of            1563–1579, https://doi.org/10.5194/hess-11-1563-2007, 2007.
crops.                                                                 Alshehhi, R., Marpu, P. R., Woon, W., and Dalla Maru, M.: Si-
                                                                         multaneous extraction of roads and buildings in remote sens-
                                                                         ing imagery with convolutional neural networks, ISPRS J. Pho-
Data availability. No data sets were used in this article.               togramm., 130, 139–149, 2017.
                                                                       Amadio, M., Scorzini, A. R., Carisi, F., Essenfelder, A. H.,
                                                                         Domeneghetti, A., Mysiak, J., and Castellarin, A.: Test-
Author contributions. This article was the result of an intensive 2-     ing empirical and synthetic flood damage models: the
week-long collaboration during the UR Field Lab in Chiang Mai            case of Italy, Nat. Hazards Earth Syst. Sci., 19, 661–678,
of June 2019. All authors attended this field lab, where the outline     https://doi.org/10.5194/nhess-19-661-2019, 2019.
and structure of the paper were agreed upon. DW and AC collected       Ames, M. G.: Deconstructing the algorithmic sublime, Big Data
and edited the input from all authors. Specific attention was given      & Society, 5, 1–4, https://doi.org/10.1177/2053951718779194,
to each section by the following authors; Sect. 1 – DW, AC, MB           2018.
and DL; Sect. 2 – DW, AC, MB, AB, EH, GMS, LR, GM; Sect. 3             Annis, A. and Nardi, F.: Integrating VGI and 2D hydraulic models
– DW, AC, RS, DL. The authors’ comments in the discussion were           into a data assimilation framework for real time flood forecast-
written by DW, AC and DL. All authors reviewed the paper before          ing and mapping, Geo-spatial Information Science, 22, 223–236,
submission.                                                              https://doi.org/10.1080/10095020.2019.1626135, 2019.
                                                                       Assumpção, T. H., Popescu, I., Jonoski, A., and Solomatine, D. P.:
                                                                         Citizen observations contributing to flood modelling: opportu-
Competing interests. The authors declare that they have no conflict      nities and challenges, Hydrol. Earth Syst. Sci., 22, 1473–1489,
of interest.                                                             https://doi.org/10.5194/hess-22-1473-2018, 2018.
                                                                       Bischke, B., Helber, P., Folz, J., Borth, D., and Dengel, A.: Multi-
                                                                         Task Learning for Segmentation of Building Footprints with
                                                                         Deep Neural Networks, available at: https://arxiv.org/abs/1709.
Acknowledgements. This invited perspective paper benefited from
                                                                         05932 (last access: 28 April 2020), 2017.
the unique intellectual environment created and facilitated through
                                                                       Bishop, C. M.: Pattern Recognition and Machine Learning,
the Understanding Risk Field Lab on urban flooding in Chiang
                                                                         Springer, Cambridge, UK, ISBN 978-0-387-31073-2, 2006.
Mai, Thailand, in June 2019 (https://urfieldlab.com/, last access:
                                                                       Bouwer, L. M., Haasnoot, M., Wagenaar, D., and Roscoe, K.: As-
23 April 2020). We would like to thank the organizers of the event
                                                                         sessment of alternative flood mitigation strategies for the C-
(Robert Soden, David Lallemant, Perrine Hamel, Katherine Barnes
                                                                         7 Basin in Miami, Florida, Deltares, Delft, the Netherlands,
and Giuseppe Molinario) for providing the unique setting to make
                                                                         1230718, 2017.
this paper possible. We would also like to thank the other partici-
                                                                       Breiman, L., Friedman, J. H., Olshen, R. A., and Stone, C. J.: Clas-
pants, who provided valuable input on this paper by participating in
                                                                         sification and regression trees, Wadsworth & Brooks/Cole Ad-
some of the discussions, namely Gautam Dadhich, Carmen Acosta,
                                                                         vanced Books & Software, Monterey, CA, USA, ISBN 978-0-
Maricar Rabonza, Pamela Cajilig, Rahul Sharma and Wahaj Habib.
                                                                         412-04841-8, 1984.
Furthermore, we would like to thank the editor (Heidi Kreibich)
                                                                       Burton, H. V., Miles, S. B., and Kang, H.: Integrating Performance-
and the reviewers (Fernando Nardi and an anonymous reviewer) for
                                                                         Based Engineering and Urban Simulation to Model Post-
their contributions to the paper.
                                                                         Earthquake Housing Recovery, Earthq. Spectra, 34, 1763–1785,
                                                                         https://doi.org/10.1193/041017EQS067M, 2018.
                                                                       Cammerer, H., Thieken, A. H., and Lammel, J.: Adaptabil-
                                                                         ity and transferability of flood loss functions in residen-

www.nat-hazards-earth-syst-sci.net/20/1149/2020/                              Nat. Hazards Earth Syst. Sci., 20, 1149–1161, 2020
1158       D. Wagenaar et al.: Invited perspectives: How machine learning will change flood risk and impact assessment

   tial areas, Nat. Hazards Earth Syst. Sci., 13, 3063–3081,                 tion mapping, Nat. Hazards Earth Syst. Sci., 15, 2725–2738,
   https://doi.org/10.5194/nhess-13-3063-2013, 2013.                         https://doi.org/10.5194/nhess-15-2725-2015, 2015.
Campolo, M., Soldati, A., and Andreussi, P.: Artifi-                      Gao, X., Klaiber, C., Patel, D., and Underwood, J.: AI
   cial neural network approach to flood forecasting                         is supercharging the creation of maps around the
   in the River Arno, Hydrolog. Sci. J., 48, 381–398,                        world, Tech@Facebook, available at: https://tech.fb.com/
   https://doi.org/10.1623/hysj.48.3.381.45286, 2003.                        ai-is-supercharging-the-creation-of-maps-around-the-world/,
Carisi, F., Schröter, K., Domeneghetti, A., Kreibich, H., and Castel-        last access: 21 August 2019.
   larin, A.: Development and assessment of uni- and multivari-           Gauss, C. F.: Theoria Motus Corporum Coelestium in Sectionibus
   able flood loss models for Emilia-Romagna (Italy), Nat. Hazards           Conicis Solem Ambientium, sumtibus Perthes, F. and Besser,
   Earth Syst. Sci., 18, 2057–2079, https://doi.org/10.5194/nhess-           I. H., Hamburg, Germany, https://doi.org/10.3931/e-rara-522,
   18-2057-2018, 2018.                                                       1809.
Carvajal, T. M., Viacrusis, K. M., Hernandez, L. F. T., Ho, H. T.,        GFDRR: Machine Learning for Disaster Risk Management, GF-
   Amalin, D. M., and Watanabe, K.: Machine learning methods                 DRR, Washington, D.C., USA, 2018.
   reveal the temporal pattern of dengue incidence using meteoro-         Ghalkhani, H., Golian, S., Saghafian, B., Farokhnia, A., and
   logical factors in metropolitan Manila, Philippines, BMC Infect.          Shamseldin, A.: Application of surrogate artificial intelligent
   Dis., 18, p. 183, 2018.                                                   models for real-time flood routing, Water Environ. J., 27,
Castelletti, A., Galelli, S., Restelli, M., and Soncini-Sessa,               https://doi.org/10.1111/j.1747-6593.2012.00344.x, 2013.
   R.: Tree-based reinforcement learning for optimal wa-                  Giacinto, G. and Roli, F.: Design of effective neural network en-
   ter reservoir operation, Water Resour. Res., 46, W09507,                  sembles for image classification purposes, Image Vis. Comput.,
   https://doi.org/10.1029/2009WR008898, 2010.                               19, 699–707, https://doi.org/10.1016/S0262-8856(01)00045-2,
Chang, L., Amin, M. Z., Yang, S. N., and Chang, F.: Building ANN-            2001.
   Based Regional Multi-Step-Ahead Flood Inundation Forecast              Goldblatt, R., You, W., Hanson, G., and Khandelwal, A.: Detect-
   Models, Water, 10, 1283, https://doi.org/10.3390/w10091283,               ing the boundaries of urban areas in india: A dataset for pixel-
   2018.                                                                     based image classification in google earth engine, Remote Sens.,
Chinh, D., Gain, A., Dung, N., Haase, D., and Kreibich, H.: Multi-           8, 634, https://doi.org/10.3390/rs8080634, 2016.
   Variate Analyses of Flood Loss in Can Tho City, Mekong Delta,          Goldblatt, R., Stuhlmacher, M. F., Tellman, B., Clinton, N., Hanson,
   Water, 8, 6, https://doi.org/10.3390/w8010006, 2015.                      G., Georgescu, M., Wang, C., Serrano-Candela, F., Khandelwal,
Chojaczyk, A., Teixeira, A. P., Neves, L. C., Cardoso, J. B., and            A. K., Cheng, W., and Balling, R.: Using Landsat and nighttime
   Guedes Soares C.: Review and application of Artificial Neural             lights for supervised pixel-based image classification of urban
   Networks models in reliability analysis of steel structures, Struct.      land cover, Remote Sens. Environ., 205, 253–275, 2018.
   Saf., 52, 78–89, 2015.                                                 Heermann, P. D. and Khazenie, N.: Classification of multispectral
Corzo, P. G. A. and Solomatine, D.: Comparative analysis of con-             remote sensing data using a back-propagation neural network,
   ceptual models with error correction, artificial neural networks          IEEE T. Geosci. Remote, 30, 81–88, 1992.
   and committee models, EGU General Assembly 2014, 27 April–             Ivers, L. C. and Ryan, E. T.: Infectious diseases of severe weather-
   2 May 2014, Vienna, Austria, 2014.                                        related and flood-related natural disasters, Curr. Opin. Infect.
Coughlan de Perez, E., van den Hurk, B., van Aalst, M. K., Amuron,           Dis., 19, 408–414, 2006.
   I., Bamanya, D., Hauser, T., Jongma, B., Lopez, A., Mason, S.,         Jonkman, S. N.: Global Perspectives on Loss of Human Life Caused
   Mendler de Suarez, J., Pappenberger, F., Rueth, A., Stephens,             by Floods, Nat. Hazards, 34, 151–175, 2005.
   E., Suarez, P., Wagemaker, J., and Zsoter, E.: Action-based flood      Jonkman, S. N., Voortman, H. G., Klerk, W. J., and van Vuren,
   forecasting for triggering humanitarian action, Hydrol. Earth             S.: Developments in the management of flood defences and hy-
   Syst. Sci., 20, 3549–3560, https://doi.org/10.5194/hess-20-3549-          draulic infrastructure in the Netherlands, Struct. Infrastruct. Eng.,
   2016, 2016.                                                               14, 895–910, 2018.
Curran, A., de Bruijn, K. M., Klerk, W. J., and Kok, M.:                  Kang, H., Burton, H., and Miao, H.: Replicating the Re-
   Large Scale Flood Hazard Analysis by Including Defence                    covery following the 2014 South Napa Earthquake using
   Failures on the Dutch River System, Water, 11, 1732,                      Stochastic Process Models, Earthq. Spectra, 34, 1247–1266,
   https://doi.org/10.3390/w11081732, 2019.                                  https://doi.org/10.1193/012917EQS020M, 2018.
De Waal, J. P.: Basisrapport WBI 2017, Deltares 1230086-002,              Keyes, O.: The misgendering machines: Trans/HCI impli-
   Delft, the Netherlands, 2016.                                             cations of automatic gender recognition, Proceedings
Dibike, Y. B. and Solomatine, D. P.: River flow forecasting using ar-        of the ACM on Human-Computer Interaction, 2, 88,
   tificial neural networks, Phys. Chem. Earth Pt. B, 26, 1–7, 2001.         https://doi.org/10.1145/3274357, 2018.
Eilander, D., Trambauer, P., Wagemaker, J., and Van Loenen, A.:           Khan, A., Khan, H., and Vasilescu, L.: Disaster Management CY-
   Harvesting social media for generation of near real-time flood            CLE – a theoretical approach, Management and Marketing Jour-
   maps, 12th International Conference on Hydroinformatics, HIC,             nal, 6, 43–50, 2008.
   21 August 2016, Incheon, South Korea, 2016.                            Kind, J., Botzen, W. J., and Aerts, C. J. H.: Accounting for risk
Eubanks, V.: Automating inequality: How high-tech tools profile,             aversion, income distribution and social welfare in cost-benefit
   police, and punish the poor, St. Martin’s Press, New York, USA,           analysis for flood risk management, WIREs Clim. Change2016,
   2018.                                                                     8, e446, https://doi.org/10.1002/wcc.446, 2016.
Fohringer, J., Dransch, D., Kreibich, H., and Schröter, K.: So-           Kingston, G. B., Rajabalinejad, M. Gouldby, B. P., and Van Gelder,
   cial media as an information source for rapid flood inunda-               P. H. A. J. M: Computational intelligence methods for the ef-

Nat. Hazards Earth Syst. Sci., 20, 1149–1161, 2020                                 www.nat-hazards-earth-syst-sci.net/20/1149/2020/
You can also read