18 December 2023
By Giulia Cirri
Last time in the AI for Climate Change series we wrote about the use of AI for flood adaptation plans and disaster relief. We have discussed devastating impacts of floods, from their immediate tragedy to the long and cumbersome process of recovery and rebuilding. Adaptation plans and disaster relief are fundamental to mitigate the impact of floods. Still, as climate change will only increase floods’ likelihood and intensity, I wanted to understand whether AI could be used in the immediate proximity to floods, to help predict floods, inform the local population and help the local community to take precautions, up to, for instance, evacuating the area. Turns out, it can. Today we are talking about AI for early warning systems.
The UN defines an early warning system as an “adaptive measure for climate change, using integrated communication systems to help communities prepare for hazardous climate-related events”. In short, early warning systems alert the population that hazardous weather is coming, allowing individuals, communities, and decision makers to take adequate measures and, if necessary, evacuate the affected areas.
Early Warning Systems (EWS) mainly consist of four parts:
If you’re familiar with AI capabilities, you might see that points 1 and 2 look like fertile ground for AI applications. Assessing large volumes of data, identifying relationships, and making predictions that can be continuously updated are some of the strong suites of AI. Before looking at the use of AI, though, one might ask: are early warning systems actually useful?
The answer, according to the World Meteorological Organisation (WMO), is yes. Early warning systems are a critical life-saving tool for floods, droughts, storms, bushfires and other hazards: we can see it in the data. Over the past five decades, recorded economic losses linked to extreme hydro-meteorological events have increased nearly 50 times, but the global loss of life has decreased significantly, by a factor of about 10. This difference can be attributed to improvements in monitoring and forecasting, and emergency preparedness.
Moreover, early warning systems are widely regarded as the “low-hanging fruit” for climate change adaptation because they are a relatively cheap and effective way of protecting people and assets from hazardous climate events. This includes floods, but also storms, heatwaves and tsunami. EWS are also economically convenient: the WMO also reports that Early Warning Systems provide more than a tenfold return on investment, and just 24 hours’ notice of an impending hazardous event can cut the ensuing damage by 30 per cent.
Nevertheless, currently, one third of the world is still not covered by early warning systems. Most of the uncovered areas are in the Global South and small island countries, which are also the areas more likely to incur the harshest impacts of climate change.
To improve the situation, in March 2022 the UN launched the “Early Warnings for All ” initiative, led by the WMO. The initiative sets the ambitious goal to have every person on the planet protected by early warning systems within 5 years. The UN has identified Artificial Intelligence as a key technology to reach this goal.
Flood predictions require several types of data. For instance, the amount of rainfall, and the rate of change in river stage. Specific information about the type of storm such as duration, intensity and areal extent can be used to determine the severity of the possible flooding. Data on a river’s basin, such as soil-moisture, topography, vegetation cover, and impermeable land area help to predict how extensive and damaging a flood might become. The list can be really long, and that’s before even assessing the impact of the flood on the specific human-modified environment.
This leaves open large opportunities for AI, with its capacity to process large amounts of data quickly and efficiently, identifying trends and relationships that might otherwise be hidden under the overwhelming volume of information.
Several big IT companies, including Microsoft, Google, Amazon, Meta and Alibaba have expressed interest in contributing to Early Warnings for all initiatives. Not all of them are necessarily planning to contribute through their AI capabilities. But Google definitely is.
Google Research’s FloodHub model is the most talked-about flood prediction AI-based model. FloodHub uses AI to forecast river floods up to 7 days in advance thorough publicly available data sources. Published research states that the model outperforms the state of the art global hydrology models, the European Union’s Copernicus Emergency Management Service Global Flood Awareness System, across all continents, lead times, and return periods.
FloodHub currently covers river basins in 80 countries. It joins two models to predict floods: the first model, the Hydrologic Model, identifies whether a river is expected to flood by processing weather and basin data, producing a forecast for the water level in the river in the following days. The second model, the Inundation Model, simulates the behaviour of the water as it moves across the floodplain based on the hydrology forecast and satellite imagery of the area. Combined, this knowledge allows the system to understand which areas are going to be affected, and how high the water level is expected to be.
While Google’s model is perhaps receiving the most traction, Google is not the only organisation looking to employ AI in flood prediction. For instance, researchers at the University of Sheffield collaborated with Environmental Monitoring Solutions, a Sheffield SME, to develop a new, award-winning technology that uses AI to prevent flooding in urban areas. At Lancaster University, researchers are collaborating with the European Space Agency’s Φ-lab to develop more accurate and explainable predictions of flooding using AI and satellite data.
As often when talking about AI, one big concern is data availability. Google reported how it relies on real-time data for only a few locations, which limits its ability to accurately predict floods. Moreover, there is a relationship between lower GDP, vulnerability to floods, and data gaps or deficiencies. Countries with lower GDP are likely to be more vulnerable to floods, but also to have more limited data availability, which complicates the use of accurate early warning systems.
As we have mentioned, different landscapes react differently to extreme events. The type of soil and the geomorphology of an area, the degree and type of vegetation or land cover, are all aspects that can affect the landscape’s reaction to heavy rainfall, which can make a particular terrain prone to flash floods. As Anthony Rea, Director of WMO’s Infrastructure Department, has said “We don’t need just to understand what weather is going to do, but what the weather is going to do to people and to the environment”.
AI can be used to bring the prediction process beyond weather modelling. An expert we met in previous work at OI once defined this new stage as “impact systems modelling”: using AI to examine how extreme events map onto local ecosystems, and how ecosystems will behave once impacted or affected by a particular extreme event.
The reaction of a particular ecosystem, or landscape, to an extreme event is too variable to be captured by traditional physics-based models. AI can provide an alternative answer, combining different data sources to identify hidden patterns: satellite remote sensing of landscape composition and digital elevation models; in-situ data on the landscape’s composition, for instance through sensors in soil; and historic data on the impacts of extreme events.
This data could even be combined with socio-economic and demographic data to understand the social impact of extreme weather events. For instance, understanding the impact on areas with more fragile population groups, such as the elderly and children. This information might be useful for several other climate-related hazards, such as heatwaves. Being able to anticipate the impact of heatwaves and map them onto different layers of the population could support targeted relief measures.
These uses of AI would require bringing together data of different types and levels of granularity, which historically has posed an important challenge to the application of AI. Moreover, these sorts of systems, especially when using socio-demographic data, would raise several concerns on ethics and data privacy. Biassed data, lack of transparency, inadequate representation and ethical use of data are still fundamental problems at the centre of the uses of AI. In terms of climate adaptation and mitigation, these issues connect to a broader conversation about climate justice and representation. In the next iteration of the AI for Climate Change series, we will cover some of these issues.
This article concludes our first round of the AI for Climate Change series. Thank you for being a part of it! We hope this series will be a useful resource for anyone interested in the connection between AI and climate change. We would love for you to be part of the conversation. If you have insights, ideas, or want to contribute to our series, please reach out to us at firstname.lastname@example.org.