Developing a new AI early warning system for flooding
Lancaster University researchers are developing new artificial intelligence systems that could help to predict and warn of flooding.
Professor Plamen Angelov leads a collaborative project with the European Space Agency’s Φ-lab to develop a new AI system that will be able to efficiently analyse and interpret streams of satellite imagery and other data in real time.
The new system would be able to improve flood alerts and rescue planning.
In previous work Professor Angelov, Chair in Intelligent Systems at Lancaster University’s School of Computing and Communications, has developed an ‘explainable’ approach to developing deep learning AI systems, called xDNN. This explainable approach helps to overcome a significant challenge in AI called the ’black box’ problem – where humans cannot understand why an AI system makes a particular decision.
By being able to more quickly, and accurately, detect floods in real time from images and other data captured by satellites we can help improve alert systems and aid rescue planning, as well as the assessment of potential damage from floods.
Professor Plamen Angelov
This three-year research project, called ‘Towards explainable AI for Earth Observation (AI4EO): a new frontier to gain trust into the AI’, will build upon, develop further and apply Professor Angelov’s xDNN system to images and other data captured by the European Space Agency’s Sentinel satellite programme.
The methods and algorithms developed within this project will process the images and other data faster and more efficiently than is currently achievable with added explainability and ability to ‘parallelise’ – splitting and running the algorithm simultaneously on multiple computers to get faster results.
Historically, automatic identification of water from satellite images involved a lot of manual steps and assumptions. It also often lacked robustness and accuracy in early flood detection.
More recent deep learning AI systems have achieved high accuracies and replaced the manual steps – however they suffer from the black-box problem where their complex architecture makes them computational and energy demanding, opaque and impossible to explain how decisions are made.
This research project into a new type of deep learning aims to develop more accurate and explainable predictions of flooding. The system should be faster to train, learn continuously and require less computational and electric power, which is very important especially for installations on satellites where the resources are scarce.
Professor Angelov, who is Director of the Lancaster Intelligent, Robotic and Autonomous Systems Centre, said: “All of us are aware that weather forecasting has improved over the years and that we are now able to receive updated weather alerts for rain on our phones. Flooding has huge economic and social impacts and often people only have short timeframes to react.
“By being able to more quickly, and accurately, detect floods in real time from images and other data captured by satellites we can help improve alert systems and aid rescue planning, as well as the assessment of potential damage from floods. This has potential for significant societal benefit.”
Is this page useful?Yes No Report an issue on this page
Thank you. If you have 2 minutes, we would benefit from additional feedback (link opens in a new window).