Floods are one of the most devastating natural disasters, accounting for the highest insured and uninsured losses annually, as well as costing many lives. With climate change possibly intensifying the hydrological cycle, the frequency and magnitude of extreme hydro-meteorological events, and therefore the risk of floods, are projected to continue to increase. This will be of devastating consequences, as it would put a greater strain on humanitarian response efforts and future financial risk of the global (re)insurance market.
Earth Observation data-based solutions currently provide a more advanced alternative to traditional ground-based flood monitoring methods or computer models, namely the ability to cover wider areas, frequent revisit times, abundant open access data and long historic image archives. However, there are still important challenges left unaddressed that compromise the quality and reliability of the data, such as the persistent cloud cover during floods, latency issues and the problem of getting abundant high-definition images under less favorable weather conditions and at night.
FloodSENS overcomes these issues by developing a flood mapping application that is capable of integrating a wider range of EO datasets and derivative data from digital elevation models using Machine Learning. This novel application being developed to market, seeks to efficiently reconstruct flooded areas under partial cloud cover in satellite images, thus creating far more reliable flood risk assessments and flood mapping during emergencies.
FloodSENS is especially important for disaster response agencies at regional, national, and international level, who are keen to utilize the proliferation of open satellite data for flood mapping during emergencies. Additionally, in the insurance and re-insurance markets, stakeholders are interested in EO data to map the flood hazard of high-impact events and on a historical basis to understand risk exposure and the changing nature of it.
In the case of the flood disaster response markets, customers often have to deal with optical satellite imagery that is partly covered by clouds during flood events, where the data resolution is too low (>30 m pixel) )to allow for local scale flood analysis and they often lack resources to deal with complex EO image analysis. This inevitably compromises the humanitarian relief efforts as it leads to incomplete estimation of flood areas and thus misrepresentation of the real impact of the flood. The (re)insurance market struggles mainly with assessing the extent of flooding for high impact events, as well as understanding the potential flood risk exposure at a local-scale. This is due to many factors, including those referred to above, along with high costs in conducting on-site inspections, using incomplete EO data archives to build historical records and prediction models, and often lack of EO specialists.
FloodSENS is targeting both the humanitarian and disaster relief organizations, as well as the global (re)insurance market, with the aim of having the application work in diverse environments worldwide.
At an initial stage, as representatives of their respective customer markets, FloodSENS will have as partners and testing customers the United Nations World Food Programme (UN WFP), the National Disasters Management Institute of Mozambique (INGC) and Willis Re (re)insurance broker through Willis Towers Watson (WTW).
FloodSENS consists of a fully automated Machine Learning-based flood mapping algorithm, whose main characteristics include:
The schematic below, illustrates the FloodSENS algorithm structure, the overall architecture, and the key submodules.
The main elements that make up the full package of the development of FloodSENS are:
FloodSENS seeks to resolve some of the biggest challenges in utilizing EO data for flood mapping, namely the cloud cover in images and low resolution, typically much coarser than 30 m in pixel size for current global operational maps. In fact, no mapping app exists on the market that maps floods consistently in different biomes at the global level , while overcoming the major limitations of optical EO imagery during floods.
This is exactly what the innovation of FloodSENS is addressing by using a machine learning algorithm to reconstruct flooding under clouds in partially clouded optical satellite images. The final map output will not only have an accurate reconstructed flood map, but will produce maps at consistent accuracy levels across different biomes.
The added value of FloodSENS centers around two major innovations:
In fact, for both Europe and the wider world, future Earth Observation-based and ML-powered apps would add considerable value to the existing products of the free Copernicus Emergency Management Service (EMS) and beyond. RSS-Hydro’s FloodSENS will place itself at the intersection of these two fields (EO technologies and AI/ML application tools) being at the forefront of future EO-enabled innovative solutions, to make a difference in allowing a much more effective disaster response.
Development work and iterative testing on several testcases, located in different parts of the globe (e.g. Mozambique, Balaton region, Greece, Bremen) are ongoing. RSS-Hydro uses Open Street Map (OSM) data as reference data, and also selected official Sentinel-1 SAR-based flood maps from the Copernicus EMS. The team has now tested a number of use cases and have a robust plan moving forward. Furthermore, RSS-Hydro has completed the automation of all data pre-processing and training steps and is now also able to fully utilize the Azure platform for development and inference. We have built a first deployable FloodSENS version for WASDI, which is not publicly released yet. WASDI is the online cloud computing platform that allows location-specific inference as well as validation by our end-user partners. The next steps now are to test the FloodSENS architecture for transferability and to continue working on customer-led flood use cases. Together with our partner end-users, we are developing user specific demos to get a better understanding of possible sector-specific added values of FloodSENS.
Bertrand.Le.Saux@esa.int