FloodSENS: Smart Sensing of Floods Under Clouds

Description

Floods are the most common type of disaster in the 21st century, with a single flood event potentially affecting hundreds or thousands of people. Current disaster response efforts rely on ground monitoring methods which are scarcely available and sometimes uncertain in developing countries, a challenge that the local response teams need to face. Uncertain information can divert the response teams’ efforts to locations that are not or only minimally affected while missing information increases the risk of critically impacted locations not being identified at all.

To enable an efficient and effective response, reliable information for large spatial areas is required. Remote sensing, and particularly Earth Observation (EO), can be used since it offers many advantages over traditional ground-based monitoring methods or computer models: large spatial coverage; frequent revisit times; abundant open access data, and long historic image archives, particularly in the case of optical imagery. However, uptake has been slow, due to two main limitations: interpretability of data and missing data due to cloud cover.

An ever-growing variety of instruments are being deployed in orbit providing an extraordinary opportunity to guide response activities on ground. For flood mapping, two instruments can be used, namely Synthetic Aperture Radar (SAR) and optical sensors. SAR, introduced in 2014 with the launch of Sentinel-1A, has the major advantage of detecting floods through cloud coverage but is hard to interpret without extensive expert knowledge and historical archives are limited. Optical Imagery is available for free for a much longer time with the introduction of Landsat in 1975 and an abundant quantity of historical data is available enabling the study of many past events. During floods persistent cloud cover is frequent thus, hiding vital information in optical imagery. Additionally, floods manifest differently depending on the terrain and soil type making some unprocessed optical images tricky to interpret.

The FloodSENS project aims to create an algorithm that efficiently reconstructs flooded areas under partial cloud cover in optical satellite images, using Machine Learning (ML) and auxiliary high-resolution data from digital elevation models and water flow algorithms. 

Supervised Machine Learning Methods essentially learn correlations between their input and the expected output through provided example. For FloodSENS, the most important input is the optical satellite image. By providing examples of known floods from the past the algorithm is expected to learn about the typical colour and shape of a flooded area. To succeed, the flooded areas need to be visible because the colour will be the key information the algorithm has to work with. As soon as clouds are present this simplistic approach will break down. Luckily a global digital elevation model (DEM) is freely available, describing the height above sea level of the terrain. The location of floods is driven by the terrain’s elevation and slope and water will always flow downhill and accumulate in depressions or rivers that might overflow. With a DEM the Machine Learning algorithm can start spotting correlations between low laying areas, the slope and the flood extent. Additional data, such as soil type and land use, can be introduced to train the Machine Learning algorithm even further. We adopt an approach to start with minimal information and slowly build up the complexity of the Machine Learning algorithm to meet important accuracy requirements.

FloodSENS is especially important for disaster response agencies at regional, national, and international level, who are keen to leverage 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.

Partners

This interdisciplinary project is carried out by RSS-Hydro in partnership with ESA’s InCubed program. Other partners? Should I mention the names of people working on this?

Location

Duration

18 months.

Type

Innovative project supported by the ESA’s InCubed program.

Impact

The added value of FloodSENS centers around two innovations:

  • The ability to reconstruct flood area under clouds in optical satellite images. This allows to valorize flood images with high cloud cover and can identify potentially missing flood areas. It also builds a more accurate and reliable historical record of flooded areas
  • Map floods in free optical satellite images at higher resolution using a trained Machine Learning (ML) model, which can produce more accurate and consistent flood maps. This in turn leads to a better estimation of impacted people and exposed assets. The resulting data are also much easier to interpret by non-experts

Publications