CSS Analyses in Security Policy CSS

ETH Zurich
No. 260, April 2020

Integrating AI into
Civil Protection
Whether by predicting the global spread of the novel coronavirus
originating from China or detecting wildfires in California, the use
of artificial intelligence in civil protection promises to improve the
prevention of, response to, and recovery from disasters. However,
getting there requires rethinking data silos and assessing high-risk
applications.

By Kevin Kohler and

Benjamin Scharte

Artificial Intelligence (AI) is an umbrella

term for a field of research and applica-
tions. Today, the term is mostly used to re-
fer to the subfield of machine learning,
which describes a set of algorithms that
rely on statistical learning. In recent years,
advances in computing power, data avail-
ability, and scalable algorithms have led to
a series of breakthroughs and AI has be-
come an enabling technology with applica-
tions across all sectors. The focus of this
analysis is the use of AI in civil protection.
Civil protection encompasses the manage-
ment of the whole spectrum of collective
risks stemming from environmental, soci-
etal, and technical hazards, such as extreme
weather events, pandemics, terrorist attacks
or severe industrial accidents, in order to In some countries, police forces use unmanned aerial vehicles in order to ensure safety compliance with
limit damage and protect the population measures taken by governments to prevent the spread of the coronavirus. Irakli Gedenidze / Reuters
and its livelihoods.

Successful applications of AI to civil pro-

tection can help to reduce deaths and eco-
nomic losses from disasters and are, there-
fore, part of the “AI for Good” paradigm,
which aims to leverage AI to achieve the
UN’s Sustainable Development Goals. In
September 2020, the International Tele-
communications Union plans to host the
fourth AI for Good: Global Summit in Ge-
neva supported by the Swiss government.
At the same time, the fact that wrong deci-
sions in civil protection can potentially cost
lives, accentuates the need for robustness,
security, fairness, and interpretability of al-
gorithmic decision-support and decision-
making.
Broad Application Spectrum
Due to the enabling nature of AI and the
breadth of civil protection, there are poten-
tial AI uses across numerous hazards and at
all stages of the disaster management cycle.
AI models are used for dynamic risk analy-
ses that can help to identify vulnerabilities,
detect hazards early on, and predict their
development, for example. Instances of this
are regularly updated local flood and land-
slide forecasts that enable the distribution
of location-based and timely safety warn-
ings. AI can also be applied in the manage-
ment of critical infrastructures. Smart elec-
tricity grids, in particular, include applica-
tions for the distributed control and
optimization in microgrids, the classifica-

© 2020 Center for Security Studies (CSS), ETH Zurich 1

CSS Analyses in Security Policy  No. 260, April 2020

tion of the types and severity of grid break-

AI in the Swiss Federal Office for Civil Protection (FOCP)
downs, as well as forecasts of electricity de-
mand, the electricity price, and the electric- Microbiology laboratories have been among the first to adapt AI into their workflow by applying
ity output of photovoltaic and wind energy. algorithms to tasks such as chromogenic detection or colony counting. The labs today need to
provide timely results with limited resources and growing challenges to society as a whole, such as
increased mobility of emerging pathogens and antibiotic-resistant bacteria. Spiez Laboratory, the
Moreover, the combination of ubiquitous Federal Institute for NBC Protection within the FOCP, currently uses AI in certain areas of analysis,
cameras with AI-enabled video analytics is detection and diagnostics, such as the selection of biomarkers in mass spectra, the differentiation
rapidly increasing the legibility of crowds and identification of pathogens or the identification of specific species of vectors in surveillance
and public spaces. While this raises serious projects. In the medium term, the FOCP could use new possibilities for data processing for more
privacy concerns, the improved situational efficient advisory services, optimizing the emergency call acceptance during crises or increasing
the reliability and response-times of alarm-systems.
awareness has applications in the manage-
ment of terror threats, pandemics, and
mass migrations. AI can also support
search and rescue efforts in the response to
a disaster. For example, AI-enabled un-
manned systems can safely provide high- Control and Prevention (CDC) and the some Chinese cities, police forces even
resolution information of operating envi- World Health Organization (WHO). wear “smart helmets” to identify persons
ronments. At the same time, advances in However, rumors amongst Chinese doc- with fever and the app Health Code uses
natural language processing increasingly tors also spread to the West through per- the mobile location history of individuals
offer real-time translation capacities, which sonal communications and human judge- to assign green, yellow, or red risk codes to
can help in situations involving linguistic ment was still needed to assess the severity them, based on which they are allowed to
minorities, tourists, or international aid of automated alerts. enter buildings and public spaces or forced
missions. to quarantine. Given the enormous eco-
The ongoing coronavirus crisis has also led nomic costs and the drastic reduction of
It is beyond the scope of this analysis to to a flurry of new AI applications, whose individual liberties caused by an indiscrim-
provide an account of how AI is used across effectiveness is still difficult to determine at inate lockdown, it is likely that more soci-
the whole spectrum of hazards. Given the this point. For example, AI-systems are be- eties will consider some form of automated
global coronavirus pandemic and the re- ing applied to triage and diagnosis. Com- and fine-grained monitoring of significant
panies like Infervision, YITU, parts of the population. However, such
In democratic societies, and AliBaba have built AI-sys- measures can be controversial, as they dou-
tems to diagnose coronavirus ble as a means of social control and govern-
automated monitoring has disease (COVID-19) in CT ments might be tempted to prolong them
to be accompanied by a scans and thereby support over- beyond an initial state of exception. Hence,
worked medicine personnel. in democratic societies, any such program
political debate on privacy, data Babylon Health, which works has to be accompanied by a political debate
protection, and civil rights. with the National Health Ser- on privacy, data protection, and civil rights.
vice in the UK, and many other
cent devastating wildfires in California and start-ups have incorporated the diagnostic Taming Wildfires
Australia, these two types of hazards are criteria of COVID-19 into their medical In the context of wildfires, analytics of sat-
used as main examples. chatbots. However, due to a lack of accu- ellite pictures, unmanned aerial vehicles, or
rate training data on COVID-19, these webcams can help to automatically detect
Hybrid Disease Surveillance symptom checkers currently rely on hand- unplanned fires, thereby reducing the re-
The nowcasting and forecasting of trans- crafted rules rather than machine learning. sponse time and increasing the chance that
missible diseases traditionally relies on lab- Furthermore, AI is also deployed in the the fire can still be extinguished. Further-
oratory evidence, sentinel networks of doc- quest for therapeutics. Deepmind has re- more, firefighters rely on modelling tools
tors, death certificates, and medical claims. leased computational predictions of pro- such as FARSITE to predict the develop-
In 2008, Google launched Flu Trends, an tein structures related to the novel corona- ment of wildfires based on topography,
attempt to predict the spread of seasonal virus to help researchers to better weather, and fuel, which can help them to
influenza in a timelier way based on online understand it. SRI International and Itkos plan managed fires and to place fuel-free
searches. However, the data proved to con- announced a partnership to use generative fire lines that limit the extent of unplanned
tain too much meaningless or misleading modelling AI technology for automated fires. In California, the WIFIRE project by
noise, and the program ended in 2015. Hy- anti-viral drug discovery. the National Science Foundation and the
brid systems that use traditional data as San Diego Supercomputer Center has
well as natural language processing of me- Lastly, states such as China, Russia, and Is- built Firemap, a web platform for the real-
dia signals are more promising. Companies rael, have turned to their domestic surveil- time and data-driven simulation, predic-
like BlueDot and Metabiota as well as the lance systems to enforce physical distanc- tion and visualization of wildfire behavior,
HealthMap at Boston Children’s Hospital ing and trace the contacts of suspected or which is used by the Los Angeles Fire De-
offer global infectious disease surveillance confirmed COVID-19 cases. Chinese fa- partment and more than a hundred local
based on mixes of AI-filtering of news ar- cial recognition companies, such as Sense- rescue services. It is important to note that
ticles, travel patterns, and expert assess- Time and Megvii, have adapted their AI- machine learning is not the only approach
ments. All of them were able to alert their systems to identify persons despite wearing within this set of digital tools. For example,
clients to the outbreak of a novel virus in facemasks, as well as to detect persons who both NASA and the European Forest Fire
Wuhan before the US Center for Disease do not wear facemasks or have a fever. In Information System (EFFIS) rely on

© 2020 Center for Security Studies (CSS), ETH Zurich 2

CSS Analyses in Security Policy  No. 260, April 2020

handcrafted algorithms for active fire de- tems also creates the threat of skill and
Further Readings
tection from satellite data. However, AI is knowledge erosion among employees
particularly used to classify land cover, to within critical infrastructures. Particularly, Bullock, J., Luccioni, A., Pham, K., Lam, C., &
detect smoke and fire perimeter on web a reliance on AI systems for monitoring Luengo-Oroz, M. (2020). Mapping the
Landscape of Artificial Intelligence Applica-
cameras, to predict fuel build-up and dan- and regulating systems, such as the energy
tions against COVID-19.
gerous weather conditions, as well as, more grid, turns humans into more passive sys- https://arxiv.org/pdf/2003.11336.pdf
experimentally, to mine radio communica- tem elements. However, human personnel
tions and social media posts. will still need a comprehensive understand- European Commission. (2020). White Paper:
On Artificial Intelligence - A European
ing of the system for the active handling approach to excellence and trust. COM(2020)
Accentuated Risks and competent reaction to malfunctions, 65 final.
Notwithstanding the opportunities AI which is particularly relevant in the unlike-
GFDRR. (2018). Machine Learning for Disaster
provides, it also creates new or accentuates ly yet consequential scenario of a large- Risk Management. Washington, DC: GFDRR.
existing risks for society. Generic risks of scale cyberattack on critical infrastructure.
AI applications include concerns about IDAG KI. (2019). Herausforderungen der
künstlichen Intelligenz. Bericht der interde-
data privacy, the perpetuation of racial and Implementation Challenges partementalen Arbeitsgruppe “Künstliche
gender biases contained in training sets, AI is no silver bullet. The current wave of Intelligenz” an den Bundesrat.
the robust and secure performance of clas- AI relies on statistical learning from large
Simonsen, L., Gog, J., Olson, D. & Viboud, C.
sifiers, the interpretability and auditability datasets. While it has many exciting appli- (2016). Infectious Disease Surveillance in the
cations, it is also severely re- Big Data Era: Towards Faster and Locally
A lack of training data is stricted in its ability to learn ab- Relevant Systems. The Journal of Infectious
stract concepts from a small Diseases, 214(4),380 – 385.
particularly pronounced for number of examples, to under- Stanley, J. (2019). The Dawn of Robot
low probability, high impacts stand causality, to transfer Surveillance AI, Video Analytics, and Privacy.
learning between domains, to American Civil Liberties Union.
events. deal with hierarchical struc-
tures, and to exhibit common-
of the decision-logic of AI, as well as the sense reasoning. If no suitable historical
accountability and liability for algorithmic data or virtual training environments exist,
decision-making. The same risks apply to AI systems cannot be expected to perform
the use of AI in civil protection, but they well. For example, the accuracy and resolu- relevant data sets accessible to these actors.
are often pronounced more strongly due to tion of electricity grid demand forecasting Due to the heightened requirements for
the decision environment. For example, depends on the availability of granular data robustness, transparency, and resilience,
search and rescue teams operate in a high- from smart meters. Such a lack of training off-the-shelf solutions are not always a vi-
stress and high-stakes environment and, data is particularly pronounced for low able option. If they are, procurement might
therefore, require reliable and tested tech- probability, high impact events as well as still require the ability to test and under-
nologies. Many AI-enabled unmanned risks from emerging technologies. stand market solutions, such as in the fa-
systems are not yet at the technology read- cial recognition vendor tests by the US
iness level to perform well under degraded Another key challenge is data aggregation National Institute of Standards and Tech-
communication conditions and extreme and standardization. Data about hazards or nology.
weather. Hence, the human-machine civil protection services is often held in
teaming needs to be stress-tested in simu- data silos. In order to create training sets Switzerland: Stronger Together
lated disasters. with sufficient quantity and quality of data, The challenges of data availability and pub-
data needs to be shared as well as labelled lic sector AI expertize are both particularly
Moreover, the use of AI in the public sector in standardized ways. As an example of ef- pronounced in the context of decentralized
arguably comes with heightened demands forts to overcome compartmentalized data political structures. In Switzerland, civil
for transparency, the prevention of discrim- structures, the Western Fire Chiefs Asso- protection is organized between the part-
ination, and integrity, as governments de- ciation in the United States provides 29 fire ner organizations of police, fire brigades,
rive their legitimacy from serving the pub- departments serving over 20 million people health services, operators of critical infra-
lic good. Most AI applications in civil with a software platform to conduct per- structures, and civil defense. As in Germa-
protection provide services to professionals formance analysis in a common framework ny and Austria, a lot of the responsibility
rather than the broader public. Still, in cas- with a centralized and normalized data set for peacetime disaster management falls on
es where public algorithmic decision-sup- with clear workflows. The integrated data subnational levels. However, the federal
port and decision-making has considerable set is anonymized and shared through the government is responsible for research and
consequences for individuals, it is impor- cloud-based Fire Data Lab. development and in agreement with the
tant to provide comprehensible informa- cantons, it may take over the coordination
tion about their decision logic. Lastly, there is an undersupply of data sci- and, if necessary, the leadership of disaster
ence and machine learning skills. The AI management.
In addition, critical infrastructures need to talent is concentrated at big tech firms and
be resilient under exceptional circumstanc- universities, whereas governments lack the The IDAG KI report particularly high-
es. As highlighted in the December 2019 in-house capacities to build their own AI lights the need for the cross-departmental
report of the Swiss interdepartmental models. Hence, the public sector needs to recording of processes and comprehensive
working group on AI (IDAG KI), the in- collaborate with universities and the pri- data access within the federal administra-
creasing complexity and opacity of AI sys- vate sector, which might include making tion. In the context of civil protection, es-

© 2020 Center for Security Studies (CSS), ETH Zurich 3

CSS Analyses in Security Policy  No. 260, April 2020

tablishing shared and normalized data sets provision, robustness and accuracy, and hu- The global coronavirus pandemic has once
between local partner organizations, the man oversight. Whether Switzerland will more highlighted the value of international
cantons, the federal government, and inter- consider similar regulations for sensitive data sharing and scientific collaboration to
national partners is arguably even more AI applications in civil protection remains protect populations. Ad hoc cooperation
important. to be seen. efforts, such as the COVID-19 open re-
search dataset (CORD-19), can make a
The technical competence gap in the public Towards Global Data Commons difference. However, the systematic build-
administration can be filled through pool- On the international level, the last decades up of global hazard datasets before the next
ing expertize, close collaboration with uni- have seen a focus on building up compara- disaster is equally important. Hence, the
versities as well as private-public partner- ble structures and establishing a global data post-Sendai agenda for disaster risk reduc-
ships. The IDAG KI report tion should prioritize sharing data about
particularly recommends the The challenges of data availability hazards themselves as well as the public
evaluation of a competence net- policy responses to them. The initiative to
work for the application of AI and public sector AI expertize conceptualize, build, and scale global data
in the federal administration. are particularly pronounced in commons by the AI for Good community
As most research and develop- might be a vehicle for such efforts.
ment linked to civil protection decentralized political structures.
occurs in academia, this collab-
oration is crucial as well to test and expand set of disaster outcomes. Specifically, the
applications in promising areas. Once an Sendai Framework for Disaster Risk Re-
application reaches maturity, public-private duction 2015 – 2030 includes the national For more on Socio-technical
partnerships are a good alternative to off- reporting of economic losses, deaths, strat- Resilience and Disaster Preparedness,
the-shelf products. egies, and development aid related to disas- see CSS core theme page.
ters. Switzerland actively supports these ef-
Finally, safety and security are more impor- forts as a voluntary donor to the UN Office
tant than deployment speed in the context for Disaster Risk Reduction, and the Sen- Kevin Kohler is a Researcher in the Risk and
of critical infrastructure. The white paper dai monitoring is important to track prog- Resilience Team at the Center for Security Studies
on AI by the European Commission sug- ress and to nudge states into action. How- (CSS).
gests that a regulatory approach should in- ever, governments also find it difficult to
clude mandatory standards for high-risk gainfully use this data to predict how haz- Benjamin Scharte is head of the Risk and
applications with regards to training data, ards evolve or what preparations and inter- Resilience Team at the Center for Security Studies
keeping of records and data, information ventions are most effective. (CSS).

CSS Analyses is edited by the Center for Security Studies (CSS) at ETH Zurich.
Each month, two analyses are published in German, French, and English. The
CSS is a center of competence for Swiss and international security policy.
Editors: Julian Kamasa, Fabien Merz, Lisa Watanabe, Benno Zogg
Proofreading: Lisa Watanabe
Layout and graphics: Miriam Dahinden-Ganzoni
ISSN: 2296-0244; DOI: 10.3929/ethz-b-000408269
Feedback and comments: analysen@sipo.gess.ethz.ch
More issues and free online subscription:
www.css.ethz.ch/en/publications/css-analyses-in-security-policy
Most recent issues:
Ukraine: The Religious Dimension of the Conflict No. 259
The Colombian Trap: Another Partial Peace No. 258
Current dynamics of urban military operations No. 257
Space Security: The Next Decade No. 256
New Technologies for Border Controls in Europe No. 255
The Middle East and China’s Belt and Road Initiative No. 254
© 2020 Center for Security Studies (CSS), ETH Zurich 4