FLORIA: Detecting urban flooding using satellite imagery and AI

Detecting floods as early as possible is a major challenge for public authorities and affected populations alike. Early warnings help to optimize rescue operations, identify which roads are passable and which are not, and assess the budget required for rehousing, reconstruction and so on. To ensure maximum reactivity in the face of such natural disasters, the ICube laboratory's SERTIT platform offers a rapid mapping service, based on the use of space images. This is possible anywhere in the world, without having to rely on measuring equipment installed in the affected area.

Launched in 2022, with the support of Carnot TSN, the FLORIA project is fully in line with this objective. Its aim is to design an automated flood detection system for urban environments, based on specific satellite data. " We are using images provided by the Sentinel-1 mission, as part of the European Space Agency's Copernicus program," explains Ari Jeannin, radar remote sensing engineer with the ICube-SERTIT platform. " Firstly, because the images are made available free of charge. But also because of the ability of the mission's satellites to operate day and night, including in the presence of clouds. " And yet, flooding is often accompanied by thick cloud cover...

Studying signal amplitude after reflection

The FLORIA project focuses on flooding in urban environments, because satellite data analysis in this context is more complex than outside cities. " The on-board radar sends a wave towards the ground, the trajectory of which depends on the obstacles encountered ", explains the engineer. " In urban environments, the signal will bounce off the road, then buildings, before returning to the satellite. On the other hand, on a plain, only a few elements, such as the roughness of the ground or vegetation, will help to reflect the signal. But in the event of flooding, the water will act like a mirror on the ground, reflecting the entire emitted wave, which will not return to the satellite at all. As a result, in lowland areas, the study of a single image is sufficient to deduce the presence of an excessive quantity of water on the surface.

On the contrary, in the event of flooding in an urban environment, the radar wave will bounce off the water, then the buildings, and eventually return to its point of departure. In this case, the difference in signal behavior between the presence and absence of flooding is more tenuous. The only solution is to compare several images, studying the amplitude sent back to the satellite for each one. This is because water has a higher reflection coefficient than asphalt: in the event of flooding, the signal will bounce off the ground more strongly, before reaching the buildings. As a result, the satellite will receive a wave of greater amplitude than under normal circumstances.

Monitoring loss of coherence

So all you have to do is compare the amplitudes of several radar images, and you're done? " Unfortunately, it's not that simple," objects Ari Jeannin. " Amplitude increases can have several causes, such as atmospheric effects. That's why we couple this analysis with radar interferometry. " This technique involves measuring the "coherence" of one image in relation to another, by studying a second characteristic of the radar wave: its phase, which also depends on the elements present in the signal's trajectory. Coherence thus reflects the degree of similarity between phases during two satellite passes.

" In urban structures, as opposed to agricultural or largely vegetated environments, coherence is generally high, as buildings and roads tend to be fixed," notes the engineer. " As a result, we are particularly interested in areas where there is a drop in coherence between two images. This indicator, coupled with an increase in signal amplitude at the same location, is then a sign of the likely presence of urban flooding.

An analysis in three images

Ultimately, FLORIA requires three satellite images: two before the event occurs, to study amplitude and coherence in normal conditions, and a third to identify any notable differences. The solution thus remains dependent on the quantity of images supplied, and therefore on the passage of satellites over the area concerned. " However, there are more and more satellite constellation projects, which should multiply the sources of information in the years to come ", Ari Jeannin points out.

To optimize radar data processing, the project team first developed an automated system. This tool, which takes the form of a Python language library, automatically extracts information such as amplitude and coherence from the raw images supplied by Sentinel-1. The library has also been encapsulated so that it can be easily transferred from one computer to another.

Forecasting through deep learning

This automated processing results in the provision of a vast amount of data, relating to the three satellite passes. All that then remains to be done is to analyze the data and deduce the probability of flooding in urban areas. A task for which the human eye quickly shows its limitations. That's why FLORIA relies on artificial intelligence techniques, in this case deep learning.

" We have benefited from recent advances in AI," emphasizes Ari Jeannin. " In our first tests with classic machine learning, we detected a lot of false positives: areas indicated as likely to be flooded, when we knew this was impossible. This was due to the noise present in the radar images. This error rate has been greatly reduced with our deep learning algorithms , capable of taking into account the probability that a neighboring pixel is also affected. " Indeed, if the presumed flooding appears to be restricted to a narrow, isolated area, there's little chance of it actually occurring. FLORIA also takes into account the shape of the pixels supposedly affected by the incident. A star-shaped flood, for example, will be considered by the tool as unlikely.

Already up and running, FLORIA continues to learn

However, the use of deep learning algorithms is always accompanied by an indispensable training phase. Consequently, the project team had to build a relevant and sufficient database from scratch. A long-term task. " We first listed around a hundred floods, and only selected those covered by Sentinel-1," recalls Ari Jeannin. " We then relied on our expertise at SERTIT to draw up maps representing these past incidents by hand, validating our observations through research on news sites or social networks. " The engineers were then able to check the relevance of FLORIA's predictions, by comparing them with their mapped descriptions of the floods studied.

This has enabled us to validate that the demonstrator we have developed is already capable of making relevant forecasts in an urban environment. " Nevertheless, we know that our model would be more accurate with more training data," concedes Ari Jeannin. This is one of the areas of improvement envisaged by the project team.

While the tool continues to prove its worth, notably in the context of the European Union's CEMS-RM program, SERTIT intends to optimize the quality of the results it delivers. This involves regular training of the model, work on AI parameters, and the use of quantified indicators to measure the relevance of the solution and track its progress.

Another indicator for improvement is processing time. " Currently, a FLORIA extraction takes between four and seven hours," notes Ari Jeannin. " We would like to reduce this time to a maximum of two hours, in line with SERTIT's responsiveness objectives. " To achieve this, the team intends to export its tool to the University of Strasbourg's HPC(High Performance Computing) platform. With the hope that its urban flooding forecasts will soon help to come to the rescue.