Policy product: Using machine learning to mitigate climate disasters
Résumés
Le changement climatique induit par les activités humaines représente une menace pour tous avec des conséquences sur différents aspects de la vie (températures élevées, sécheresses, incendies, inondations) et la dégradation des conditions météorologiques. Face à cette situation, il importe de mettre des systèmes d’alerte précoce. Ces systèmes dont le but est d’avertir les populations des orages, inondations ou épisodes de sécheresse imminents ne sont pas anodins. Ils offrent un outil efficace qui permet de sauver des vies, de réduire les pertes économiques et préjudices causés par les dangers météorologiques, hydrologiques et climatiques2. En ce sens, le Secrétaire général des Nations Unies, António Guterres, en 2022, a lancé un appel mondial pour que tous les habitants de la planète soient protégés par un système d’alerte précoce multi-aléas (MHEWS) d’ici à 2027. Définie comme l’ensemble des algorithmes permettant d’analyser les données afin d’identifier des modèles, des tendances, des corrélations et d’autres informations pertinentes qui améliorent le processus décisionnel, l’Intelligence Artificielle (IA) est de plus en plus sollicitée dans les prédictions météorologiques. Ainsi, le présent article analyse ses différentes implications dans la prévention des risques et catastrophes naturelles à travers ses contributions aux systèmes d’alerte précoce.
Entrées d’index
Mots-clés :
changements climatiques, Intelligence artificielle, IA, alerte précoce, catastrophes naturellesPlan
Haut de pageTexte intégral
1. Objective of a multi-hazard early warning system (MHEWS)
1A multi-hazard early warning system (MHEWS) is capable of anticipating the risks associated with climatic events, monitoring these risks and taking timely action to prevent these events from turning into climatic disasters. It consists of 4 main components:
-
Knowledge of disaster risks,
-
Monitoring, forecasting and warning,
-
Distribution and communication,
-
Ability to react.
Figure 1 - https://public.wmo.int/en/media/press-release/early-warnings-all-initiative-gains-momentum

2With the frequency and scale of climatic events increasing, SMHEs are becoming more and more crucial. At the end of a year marked by some of the deadliest floods, droughts and other weather-related disasters in Africa, the UN Secretary-General pointed out that "Vulnerable communities [...] are caught unprepared by cascading weather-related disasters with no advance warning [resulting in average disaster mortality] eight times higher than in countries with high coverage".
2. Mandate for universal access to multi-hazard early warning systems
3The United Nations World Meteorological Organization (WMO) has attributed the decrease in the number of lives lost - despite the increase in the number of disasters recorded - between 1970 and 2019 to the expansion of NMHSs. At present, only half of WMO Members have a NEMS, with implementation particularly slow in developing economies, where the majority of disasters (71%) and associated deaths (91%) are concentrated. Africa, where six out of ten people are not covered by such systems, recorded 1.5 times as many deaths4.
- 5 = 3billion/800million; in terms of avoided losses
4In order to save more lives through EMS, the United Nations has called for everyone on the planet to be protected by an EMS within five years. However, the ability to translate early warnings into rapid action is often insufficient. Closer collaboration is needed at all levels, from data providers to end-users, including sectoral experts, communications institutions and the most remote and underserved people, with an estimated cost of $3.1 billion for a return of nearly 400%5 and the saving of 23,000 lives each year. The WMO is currently working on a SMHE index to monitor the deployment of coverage around the world, and the next opportunity to discuss this will be the United Nations Water Conference in 2023.
3. Multi-hazard early warning systems in Africa: Challenges
5While African nations have a key element in ensuring an adequate level of protection for their citizens, the African Multi-Hazard Early Warning and Early Action Programme, launched in 2021, aims to provide universal early warning and early action protection to every individual on the continent within five years. Three crisis units around the world monitor climate risks.
- 6 These were some key trends from a list of challenges identified from a workshop (March 2022), a Bou (...)
6For this programme to be fully effective, countries need to address several key issues6 as part of their own early warning processes.
7Firstly, knowledge of disaster risk, on which all warning systems are based, is often incomplete. The historical, temporal or spatial data collected is not always sufficient to establish complete risk profiles, particularly at the relevant administrative level, as in the case of the risk of flooding in urban areas, which is not often assessed.
8In addition, risk assessment tools do not always systematically take into account non-traditional data, such as the experiential knowledge of local leaders or that of NGOs working in the country. In some countries, knowledge of the risks associated with slow-onset phenomena, such as droughts, was more extensive than that of the risks associated with fast-onset phenomena, such as floods. In general, data relating to the phenomena themselves are more exhaustive than those relating to vulnerability and exposure to climatic hazards, which are rarely quantified. This results in lists of vulnerable people and maps of emergency resources that are often incomplete, obsolete or missing. Even when exhaustive and detailed risk assessments are available, they are not always used appropriately to inform the development of emergency plans and/or are not regularly updated when climatic events evolve and give rise to new emergency situations.
9Secondly, there are deficits in monitoring, forecasting and warning services. Geographical monitoring infrastructure, such as automatic meteorological or hydrometeorological stations, is still inadequate in many countries. What’s more, in some countries the data exchange and coordination mechanisms for real-time monitoring, forecasting and early warning are neither formalised nor automatic.
10The monitoring of certain fast-moving events, such as floods, faces greater obstacles than that of slow-moving events, such as droughts. In some countries, forecasts, particularly those relating to hydrology, are not drawn up taking their impact into account.
11Finally, shortcomings in dissemination and communication contribute to public misinformation. There are three major challenges in this area:
3.1 Understanding
12The basic terms of SMHEs can be confusing for some policy-makers, local officials and community members, making it more difficult to communicate with the general public. Language barriers and a lack of media involvement can also contribute to the communication gap.
3.2 Meeting deadlines
13In some countries, alerts are sent by email to specialists, delaying the process and posing a challenge for short-term events. Messages from ministries to districts can be inconsistent, with rural and remote areas generally the worst affected.
3.3 Taking appropriate action
14Even when alerts are issued in good time and through the appropriate channels, they are not always followed up with coordinated and integrated action as part of contingency plans. In addition, there is often no formal feedback loop to ensure that the population is targeted effectively.
15In conclusion, in terms of response capacity, miscommunication between partners and stakeholders is common. Contingency plans do not always cover the whole country and do not always take into account the diversity of vulnerabilities within the population. In addition, after-action evaluations often lack consistent and documented feedback, which hinders effective improvement.
4. Artificial intelligence: a brief explanation
16Machine learning algorithms are the driving force behind artificial intelligence systems. These algorithms analyse data to identify patterns, trends, correlations and other relevant information that improve the decision-making process. A statistical model captures the relationships between data and concepts, enabling AI systems to make predictions.
17A model can predict the next likely value in a series or determine the category to which an object or person belongs. It can also identify groupings of objects or people with significant similarities. The most advanced models are now capable of generating data such as images or text. Although machine learning (ML) is an essential component of AI, our focus in the remainder of this paper will be on it, although the term ’AI system’ will be used to encompass the whole ecosystem in which ML operates.
18Models can be divided into two categories: shallow learners (SL) and deep learners (DL). A shallow learner (SL) analyses datasets with a limited number of columns (features) and rows (instances), while a deep learner (DL) processes larger, more complex datasets using a tool called a neural network, designed to simulate brain function. SLs can rival the most sophisticated deep learning systems in performance, but the main distinction is that DLs operate more autonomously, with less human intervention.
19Deep Learning (DL) can considerably enhance the capabilities of SMES. These models are better equipped to integrate unstructured data such as text, images, audio and video files, offering a representation closer to the real complexity of the data. As a result, the knowledge acquired by DLs is often more accurate and detailed. In addition, DLs are capable of performing more advanced tasks such as image and speech recognition.
20However, DL techniques are more computationally and energy intensive than surface learning (SL) approaches. This disadvantage can be mitigated by using cloud-based solutions, where calculations are performed remotely before the results are returned. However, the ongoing need for high quality data, such as professionally written text rather than social media posts or website comments, as well as greater technical expertise, remains a challenge.
21The scale and complexity of ML models also exacerbates the ’black box’ problem, where the model becomes difficult for humans to interpret due to its increased complexity. This makes validation, evaluation and improvement of some ML systems more difficult.
5. Promoting ML in SMTPEs
22Ideally, an SMES using ML techniques would automatically collect data from a variety of sources, make probabilistic predictions early enough, feed these predictions back to decision-makers along with classified follow-up actions, and establish a feedback loop to evaluate and improve performance. However, even outside this ideal situation, ML techniques can improve individual tasks in all four components of SMES, enabling results to be achieved at a level of accuracy, speed or scale that is difficult for humans to reproduce.
5.1. Improving understanding of disaster risk
-
Studies have shown that models trained on diversified datasets perform better. AI systems facilitate the systematic collection of vast quantities of data from a variety of sources such as ground observations, street images, connected devices, participatory data, etc., in real time.
-
ML techniques such as image classification and object detection can be used to analyse aerial (satellite) images to identify features such as land cover and use, as well as built and natural infrastructure. They can also assess the condition of buildings and their construction progress.
-
ML tools bypass the need for costly field surveys or to wait for decennial censuses; satellite images of the Earth at night make it possible to monitor interactions between populations and river resources. For example, ML techniques have used satellite data, meteorological data and mobile phone metadata to identify the most deprived neighbourhoods and individuals with greater precision, surpassing other methods.
-
Risk is rarely immutable. As impacts materialise and communities adapt, their risk profiles evolve. AI systems can incorporate these changes more quickly and consistently to provide constantly updated risk profiles, using adaptive ML techniques that absorb and adjust to continuous streams of data from the environment.
-
Although a digital twin of the Earth remains a distant prospect, more modest initiatives have shown promise in creating partial ’dynamic digital replicas’, faithfully mimicking the Earth’s behaviour through ML processes. For example, a digital twin focused on hydrology would be able to identify landslide risk zones and periods, as well as flood risks.
5.2. Faster monitoring, forecasting and warning
-
ML techniques simplify the complex process of predicting outcomes in multivariate environments, making multi-hazard impact analysis more accessible.
-
Risk monitoring is sometimes split between several government agencies. By merging data from different sources, ML techniques can generate comprehensive alerts, reducing the risk of discrepancies or confusion in warnings issued by individual government agencies, which are based on limited data sets.
-
To ensure the best possible response to a risk, policy-makers need as much localised data as possible. ML-based models provide much more detailed, impact-focused predictions, enabling more efficient allocation of emergency resources, such as health services or shelter. For example, an ML algorithm can predict the damage caused by an earthquake at the level of census zones (blocks).
-
AI systems are capable of predicting second-order risks beyond the immediate danger, such as epidemics or disruptions to schooling, by exploring various scenarios.
-
The ability of AI systems to integrate unstructured data makes it easier for recipients to check that alerts have been received and understood.
5.3. Effective dissemination and communication
-
Given the diversity of climatic events and their impacts, as well as the disparities in the capacities of vulnerable populations, it is crucial to personalise warning messages. ML-based tools can facilitate this personalisation in a number of ways.
-
ML techniques can generate early warnings simultaneously in multiple vernacular languages and dialects.
-
Although the common alert protocol has standardised the dissemination of alerts via several communication channels, ML techniques make it possible to identify the most effective channels, thus avoiding overwhelming a community with redundant messages that could lead to apathy.
-
Messages can also be tailored to different demographic groups such as rural populations, women, the elderly and the disabled.
-
Finally, by directly associating messages with predicted impacts, an artificial intelligence system can determine the best resources for communities and provide more information about these resources.
5.4. Better ongoing monitoring
-
The ML techniques used in disaster simulations help to identify the best response strategies. Thanks to a more accurate prediction of the probability of second-order events, decision-makers have a wider time window in which to take beneficial measures, even beyond the immediate emergency.
-
It is crucial to monitor the status and location of vulnerable populations as risks evolve. ML techniques facilitate this monitoring by providing frequent updates on the migration of vulnerable populations, using geolocation data from mobile phones in particular, which is more reliable than other methods.
-
Up-to-date situational reports on the situation are essential to determine the most appropriate responses. ML-based tools can improve the speed and accuracy of these reports, for example by automatically identifying information gaps in affected areas and having satellites collect this data until human responders arrive on the ground.
6. Opportunities for multi-hazard early warning systems in Africa
23This table illustrates how specific ML-based solutions can help solve some of the above challenges.
Component of the multi-hazard early warning system |
||||
Knowledge of disaster risks |
Monitoring, forecasting and warning |
Distribution and communication |
Reaction capability |
|
Existing fault |
Lists of vulnerable people are often imprecise. |
Relevant or actionable alerts are not always issued in a timely manner. |
The public is poorly informed. |
Unclear processes lead to unclear answers. |
Inputs |
Mobile phone metadata, combined with traditional, structured data such as census or welfare lists, can help identify the vulnerability element of risk. |
Integrating unstructured data, such as text or voice messages from regional managers, can lead to more relevant early warnings. |
The use of unstructured data, such as audio or text files, to establish communication with indigenous communities in several vernacular languages. |
To dispel the confusion caused by an opaque process and increase accountability, it is necessary to bring to light the data hidden in the departments. |
Process |
ML techniques make it possible to identify the degree of vulnerability of a community without having to undertake costly surveys. For example, the government of Togo has used this approach to identify the most vulnerable people for targeted cash transfers. |
The evaluation of a flood forecasting model recommended collecting unstructured data in the form of WhatsApp messages from local leaders and incorporating them into alerts. |
ML techniques, such as natural language processing, can ensure fast and efficient translation between end-users and technical stakeholders, reducing confusion. |
Chatbots such as Chat-GPT can be used to create applications that answer common questions asked by disaster risk management stakeholders. |
Outputs |
The results can be used to register members of a vulnerable community or to draw up a risk map showing the location of vulnerable communities. |
The resulting alerts are more accurate and relevant to regional communities, minimising the waste of emergency resources. |
Messages are less dependent on traditional media and are more targeted at the local language or dialect of the community. |
Question-and-answer chatbots for decision-makers enable better coordination of information on emergency situations, and guarantee consistent access to information at national and sub-national levels. |
24In too many African countries, disaster management remains essentially reactive and manual. Few aspects are digitised, let alone automated, leading to gaps in service delivery. AI systems offer opportunities for innovation and, within SMES, the aim should be "continuous process improvement". This form of innovation frees up the time and cognitive energy of those involved in disaster relief management (DRM) to deal with atypical situations that require greater human intelligence or for which an automated solution is not readily available.
7. Risks associated with ML techniques in MHEWS
25Policymakers need to be aware of several major risks associated with the use of ML techniques for MNCH.
-
Collecting sufficient meaningful data for ML algorithms can be a significant challenge. All algorithms require some level of data collection, depending on the degree of adaptation to local conditions. Sometimes, even identifying the most relevant data to collect can involve costly and time-consuming trial and error. And that’s before investing in the infrastructure required for this data collection, which ranges from airborne and ground-based sensors to data transmission channels and storage capacity.
-
Collecting data on vulnerable communities in order to help them involves potential privacy risks. In the absence of strict data governance rules and enforcement, MHPSS beneficiaries can become targets of malicious actors at a time when they are most vulnerable. Sometimes the risk comes from public or private sector institutions, particularly where regulations to protect citizens’ privacy are limited. Even without malicious intent, unintentional data breaches with no recourse can be extremely damaging.
-
Making forecasts from historical data is an exercise that can be affected by biases, assumptions and out-of-date data. Climates evolve rapidly and the multiplication of extreme weather events and phenomena can compromise the reliability of forecasting models. It is therefore crucial to evaluate a model on an ongoing basis, with particular emphasis on its flexibility, predictive capacity and ability to take uncertainty into account.
-
Most ML models are designed and developed by and for non-ICMP users, which can introduce a western bias. This can affect the accuracy and reliability of the model when applied outside its original context. It is therefore necessary to exercise caution when using a model in a new environment or for a new population.
-
ML models are still far from reaching technological maturity. Deep learning techniques emerged only a decade ago, and the most recent advances, such as those in language learning models, date from the last two years. The next iteration, expected in a few months’ time, already promises improved accuracy. What’s more, there are significant gaps between what a model can achieve in a controlled laboratory environment and the challenges of the real world. Engaging with research centres to monitor these updates and assess their relevance to SMHE requires time and a commitment that policymakers may not have.
-
An ML technique without the appropriate processes and support staff will not be very effective. ML-based tools developed outside of existing DRM systems without consideration for their integration, including adequate training, may quickly become obsolete.
-
Deploying ML-based solutions without considering governance and post-implementation monitoring can lead to abandonment by potential users. Management must be actively engaged in using new solutions, including efforts for continuous improvement through feedback mechanisms. Without an appropriate level of management commitment, user resistance to change can lead to unplanned obsolescence.
8. Next step for political decision-makers
26The deployment of ML-based solutions and AI systems for climate change adaptation - of which SMES are a part - faces two main challenges:
-
Firstly, most, if not all, African countries are at the bottom of the Oxford Insights Government AI Readiness Index, which assesses, on a global scale, the readiness of governments to use AI in public services. There are several reasons for this lack of readiness, which reflects the challenges faced by many AI projects in general.
-
Secondly, because of the different priorities of policy-makers in low- and middle-income countries, adaptation to climate change, and by extension Monitoring, Early Warning and Response Systems (MERWS), is often relegated to second place in terms of resource allocation.
27Our recommendations take these challenges into account:
-
Build long-term capacity while capitalising on current opportunities;
-
Identify the most feasible actions that can bring tangible benefits.
28By following these recommendations, the players involved in natural risk management can gradually adopt solutions based on machine learning. This could eventually convince their respective governments to adopt artificial intelligence more widely and in a more integrated way.
29According to a previous TBI report, governments are encouraged to adopt a strategic approach to integrating and experimenting with artificial intelligence (AI) applications in a responsible and effective manner, taking into account current technological capabilities. This approach, represented by the pyramid below, aims both to strengthen the foundations of AI, such as skills and infrastructure, and to achieve specific objectives. Policymakers are encouraged to start at the top of the pyramid, determining the depth of their commitment based on available national capabilities, and then progress in stages. This strategy has similarities with one dimension of a two-track framework previously suggested by USAID for AI adoption in health, although the recommended actions may differ.
30Following the first step, policymakers need to formulate the question - or use case - they wish to answer. For example, "How can we deliver national flood forecast reports to regional and local officials on a daily basis?" A clearly defined question guides decision-makers on what is expected (in this case, daily flood forecasts for the appropriate administrative levels) and how to assess whether the new process represents an improvement.
31Policy makers need to identify the value chain of data, processes and personnel involved in this issue. Upstream of this issue are the departments responsible for data collection, while downstream are the end users of flood forecasts. This identifies areas where improvements can be made and ensures that the value added by AI improvements is not hampered by a wider process that is inefficient or incomplete.
32The second step is to determine whether an existing model is suitable for the intended purpose. For example, for flood forecasting, Namibia uses the United Nations International Hydrological Programme to provide daily flood risk bulletins to local communities. However, another model that predicts only one flood risk indicator, such as extreme rainfall, without taking into account other factors such as river flows or soil moisture levels, may still be useful depending on a country’s specific needs. Policymakers need to be aware of how AI-based solutions can fit into the overall SMHE process while enabling faster and more accurate predictions.
33It is difficult to estimate precisely the number of ML models available that can predict climate-related risks. For policy makers, it is of interest to identify publicly available models that are relevant to their use case (and potentially to future use cases). Several resources are available to begin this process, although most of them focus primarily on risk prediction rather than vulnerability or exposure:
-
Identify and participate in initiatives such as the African AI Innovation Council and SPACES, to establish collaboration between governments and technology companies.
-
Identify open data sources such as the European Space Agency, NASA or Google Earth Engine to see if they offer pre-trained models.
-
Explore the free online model libraries. Github is the largest online repository, but there are also TensorFlow and PyTorch.
34The models selected should be evaluated according to the following criteria4 , which countries may weight differently according to their needs and capacities:
- 7 McKinsey estimates that due to interoperability issues, 40% of AI's potential benefits may not be r (...)
Criteria |
Description |
Data |
The data used to configure the model is available free of charge, regardless of geographical location. |
Software |
The degree to which the model is based on open source code. |
IT efficiency |
The time taken to generate forecasts compared with the time taken to execute forecasts. |
Flexibility |
The model’s ability to adapt to changes and incorporate new observations through data assimilation. |
Technical expertise requirements |
The time required for local non-specialist staff to acquire the technical skills and knowledge needed to use and maintain the model, independently of its designers. |
Forecasting skills |
Forecast accuracy expressed in terms of success rate, false alarm rate and mean-based measures. |
Uncertainty |
The way in which forecast uncertainty is presented in the model results. |
Accuracy or explicability |
If the end user is to be able to explain how a final decision was reached, the model’s explicability must be taken into account. |
Assimilation |
The solution must be assimilated into existing software and MHEWS5 procedures7 . |
Infrastructure |
The IT resources required, particularly for data storage, analysis and processing, increase with the complexity of ML. However, cloud-based delivery can eliminate the need for on-site hosting for the end user. |
Representativeness |
The inclusion in the model of knowledge and data specific to the region. |
35It is unlikely that any product or solution will rely entirely on ML models. This is particularly true for risk forecasting, which has long relied on traditional meteorological modelling techniques. Combining traditional weather models with ML techniques can bring the benefits of greater accuracy and accessibility to end users. An evaluation of flood forecasting models has recommended a hybridisation of techniques. It is important that policy makers are aware of these hybrid approaches and some of the organisations developing this work: Allen Institute for AI and Artificial Intelligence for Earth Observation.
36For the third stage, if no existing ML-based solution satisfactorily addresses the use case, it may be possible to modify one. This is known as "transfer learning": a pre-trained model is identified and refined to meet the specific needs of the context using regional data. This builds on the strengths of a pre-trained model while ensuring that it is specific to the current context. The expertise and resources required to modify an existing model may vary; this decision tree (for "Developing predictive models for hazard") provides a useful check before embarking on such a project.
37The results obtained at these stages can then be used to feed into centralised monitoring processes, such as the AU’s multi-hazard early warning and early action programme for Africa or other global initiatives. These results can contribute to strengthening analysis and forecasting capacity, or be directly integrated by the country itself.
38To reach the fourth stage, if no model can be transferred or refined without additional data, a process needs to be developed for integrating these data from the sources that collect them. Satellite imagery is an example of abundant but underused data. According to climatologists, if daily rainfall or flood statistics - freely available thanks to satellites - had been monitored, more timely decisions could have been taken during the 2022 floods in Pakistan. Satellite data has been a game-changer for countries able to use this free, open data. Australia has used Sentinel images from the European Space Agency to map vegetation and thus predict bushfires. A space sector focused on the needs of low- and middle-income countries, as well as African countries, is currently booming. The Space 2022 exhibition and conference in Kenya included sessions on accessing and using Earth observation data.
39In addition to other databases, sensors and online platforms that contain relevant data, policymakers should also identify and search for digitally inaccessible but nonetheless useful ’dark’ data held by various government departments. For example, a government agency responsible for distributing welfare benefits would probably have lists of vulnerable people across the country, information that could prove useful in building a model to estimate the risk of vulnerability ahead of a climate event.
40Once identified, the data must be prepared for use in an ML process. This includes identifying and correcting errors, inconsistencies and missing values. In addition, some ML techniques require ’labelled’ data: for the model to provide an answer or prediction on a new instance of data, it must have been exposed to similar data previously. Data preparation is a fast-growing field: it is estimated that in India alone, between 500,000 and one million people are employed in data labelling. This sector is also present in Africa and is growing. Policymakers should see this as an opportunity to boost skills in the technology sector while offering appropriate guarantees to these workers. Although these jobs may eventually be automated (or performed by experts with a high level of advanced expertise), given the huge amount of unlabelled data - particularly in the field of earth observation - this sector is likely to continue to grow for at least the next decade.
41Once the data has been cleansed, it must be merged with data from different sources and in different formats to facilitate analysis and comparison. A system for storing and organising the data, such as a database or data warehouse, needs to be put in place and, finally, procedures and roles need to be established for maintaining, updating and accessing the data.
42The final step that policy makers should consider is developing a plan for collecting new data if existing data is not sufficient. In addition to identifying the type, format and source of data for the pre-defined use case and the collection channels, policy makers should consider the following:
-
Is the data of high quality? The WMO is in the process of identifying key criteria for Earth observation procedures in order to improve the reliability of risk forecasts. - Are the data reliable? Crowd-sourced data is generally less reliable than data obtained from field surveys.
-
Does the target audience understand the data? The users of these data or their derivatives must be able to interpret them correctly.
-
Do decision-makers want this data? Data and any derived products must enable users to make informed decisions.
- 8 Specifically, "Collect hazard, exposure, vulnerability data".
43A decision tree8 , together with a cost-benefit analysis, can guide investment decisions in new software and hardware, such as a network of automated weather systems and channels for transmitting data to relevant stakeholders. Finally, data must be validated to ensure accuracy and consistency.
9. Conclusion
44It is essential that every man, woman and child is covered by the protective barrier provided by FDMS. The annual benefits in terms of lives saved - 23,000 - as well as the savings made by avoiding losses and damage, which are 4 times greater than the initial investment, demonstrate the soundness of this approach both ethically and economically.
45Currently, many African countries adopt a reactive and manual approach to disaster risk management, resulting in gaps in service delivery. Policy makers should see the planning and development phase of the AU’s multi-hazard early warning and early action programme as an opportunity. Firstly, countries should exploit the benefits of the AU’s centralised monitoring effort by entering into agreements for the collection and transmission of national data, while taking advantage of its analytical and forecasting capacity. Secondly, countries should learn from this programme to improve their own early warning and early action systems, providing more accurate forecasts and more targeted messages and responses.
46These two national initiatives can benefit greatly from the use of ML techniques, paving the way for more accurate, faster and larger-scale results than human effort alone could achieve:
-
In terms of understanding disaster risk, ML techniques reinforce traditional hazard modelling and offer new ways of assessing exposure and vulnerability risks, especially in countries where data is less accessible. They also enable more effective monitoring of the rapid evolution of events.
-
In terms of monitoring, forecasting and warning, ML techniques open the way to more accurate and locally relevant predictions and tested resilience scenarios, providing disaster risk management stakeholders with a more detailed view of potential impacts.
-
When it comes to broadcasting and communication, ML techniques can provide messages that are tailored to the audience and can be exploited via the most effective channels for each population.
-
Finally, in terms of response capability, ML techniques can enhance disaster simulation tools, optimising decision-making and providing up-to-date reports on the situation, including the status and location of affected communities.
47Our recommended approach recognises and mitigates two key risks associated with the use of ML techniques for NEMS: a lack of preparedness at the national level and a lack of prioritisation of the DRM mission. By adopting our incremental and progressive approach to the adoption and assimilation of cybercrime techniques into their crisis management operational procedures, countries can take advantage of these technologies today, even if their readiness index ranking is not very high. At the same time, this approach allows countries to put in place the tools they need to improve their score. Policymakers should see each deployment as an opportunity to improve SMES and build internal AI capacity. The readiness index can be used to measure progress.
48Calls to action :
-
Identify opportunities to automate highly repetitive but time-consuming tasks within DRM operational procedures. Then critically assess the appropriate level of ML-based intervention for each task, ranging from deploying existing models to collecting new data to populate new models.
-
Establish partnerships with knowledge centres and technology companies that focus on the deployment of AI in one or more components of SMES, so as to be able to transfer rather than build from scratch. In addition, leverage the current users and customers of these organisations to better understand how they are using these tools.
-
Work with the AU’s Africa MHEWEA programme to identify opportunities to provide data to improve monitoring, especially where ML techniques can be deployed, and to obtain forecasts and alerts.
49The deployment of ML and AI systems in HMEs should be seen as an ongoing process rather than a one-off implementation. These technologies offer real improvements on the status quo. When lives and livelihoods are at stake, that is reason enough to consider them seriously.
-
Internal analysis
-
= 3 billion/800 million; in terms of losses avoided
-
These are some key trends drawn from a list of challenges identified at a workshop (March 2022), a Bournemouth University assessment (August 2021) and an UNDRR assessment (2020) which, taken together, cover five countries (Sierra Leone, Angola, Ethiopia, Tanzania, Zambia) and are not intended to be representative of the continent.
-
The first 7 criteria are taken from https://www.researchgate.net/publication/339873760_Development_and_evaluation_of_flood_forecasting_models_for_forecast-based_financing_using_a_novel_model_suitability_matrix. The others are based on internal research and analysis.
-
McKinsey estimates that due to interoperability issues, 40% of the potential benefits of AI may not be realised (Manyika et al., 2015).
-
More specifically, "to collect data on hazards, exposure and vulnerability".
Notes
1 https://wmo.int/sites/default/files/2024-06/Early%20Warnings%20for%20All_Factsheet_FR.pdf
2 https://wmo.int/sites/default/files/2024-06/Early%20Warnings%20for%20All_Factsheet_FR.pdf
3 https://wmo.int/sites/default/files/2024-06/Early%20Warnings%20for%20All_Factsheet_FR.pdf
4 Internally conducted analysis
5 = 3billion/800million; in terms of avoided losses
6 These were some key trends from a list of challenges identified from a workshop (March 2022), a Bournemouth University evaluation (August 2021) date), and a UNDRR assessment (2020) that, in aggregate, cover 5 countries (Sierra Leone, Angola, Ethiopia, Tanzania, Zambia) and is not meant to be representative of the continent.
7 McKinsey estimates that due to interoperability issues, 40% of AI's potential benefits may not be realised (Manyika et al., 2015).
8 Specifically, "Collect hazard, exposure, vulnerability data".
Haut de pageTable des illustrations
![]() |
|
---|---|
Titre | Figure 1 - https://public.wmo.int/en/media/press-release/early-warnings-all-initiative-gains-momentum |
URL | http://journals.openedition.org/ctd/docannexe/image/12855/img-1.png |
Fichier | image/png, 72k |
![]() |
|
URL | http://journals.openedition.org/ctd/docannexe/image/12855/img-2.png |
Fichier | image/png, 11k |
Pour citer cet article
Référence électronique
Sabahat Iqbal, « Policy product: Using machine learning to mitigate climate disasters », Communication, technologies et développement [En ligne], 16 | 2024, mis en ligne le 01 novembre 2024, consulté le 24 mars 2025. URL : http://journals.openedition.org/ctd/12855 ; DOI : https://doi.org/10.4000/12nfj
Haut de pageDroits d’auteur
Le texte seul est utilisable sous licence CC BY-NC-ND 4.0. Les autres éléments (illustrations, fichiers annexes importés) sont « Tous droits réservés », sauf mention contraire.
Haut de page