Détecter les inondations en milieu urbain grâce à de l’imagerie satellite et de l’IA

[ad_1]

As part of a European program, a new rapid mapping service is being tested to detect the excessive presence of water in urban areas. It was developed by a laboratory at the University of Strasbourg and works using radar satellite imagery coupled with artificial intelligence algorithms.

Copernicus Emergency Management Service (CEMS) is a program set up by the European Commission to provide precise and rapid geospatial information to all actors involved in emergency management. claims natural. Until now, in the event of flooding in urban areas and cloud cover, no service was offered to provide an overview of the extent of the damage, even though it is precisely in these inhabited areas that the population is the most vulnerable. For several months, a new tool using satellite imagery and AI has been in the testing phase to demarcate flooded areas as quickly as possible and guide emergency services. It was developed by the technological platform of the ICube laboratory at the University of Strasbourg as part of a project called Floria.

Two techniques make it possible to visualize a flood from satellite images. The first is based on optical imaging and consists of taking high-resolution photos to visualize the presence of water. Except that it does not work in the presence of clouds. These being generally very frequent during flooding due to rain, this method cannot be considered for this type of disaster. There remains then, the second technique, by radar imagingand which consists of sending an electromagnetic wave from a satellite, then analyzing the wave returned by the ground.

“As part of the Floria Project, we use this technique, explains Rémi Braun, software development manager of the ICube-SERTIT platform. It works very well in the countryside and in less urbanized areas, but much less so in the city. After hitting the road, the electromagnetic wave will bounce off buildings before returning to the satellite, which will disrupt the analysis of the signal. To deal with this difficulty, we analyze not one image, but three, that is to say two before the flood and one after. »

The analyzed images are provided by the Sentinel-1 mission

Thanks to theradar interferometry, a technique consisting of exploiting the phase difference between several radar images in order to estimate a similarity rate, scientists establish a coherence map between the images. A second map is also designed by analyzing the increase in signal intensity of the electromagnetic wave at identical locations between images. Secondly, AI algorithms are used to decipher these two maps and identify potentially flooded areas on each pixel. The model developed is of the deep learning type and works based on the architecture of neural network U-Net.

Currently, the ICube laboratory teams are relying on images provided by the Sentinel-1 mission, as part of the European Space Agency’s Copernicus program. For these images to be usable, one of the constraints is that they must be taken from the same position in space. “The advantage of Sentinel-1 is that it is the only satellite to perform systematic acquisition and that the images are available free of charge.adds Rémi Braun. On the other hand, the second satellite failed and we are awaiting the launch of a third satellite, scheduled in approximately 6 months, to improve our cartography. The time it takes for satellites to revisit the same point on earth, which depends on latitude, should be around two days in Europe. The frequency of image acquisition will then be sufficient to be able to analyze them in good conditions. »

This new digital model is in the form of a demonstrator and is currently being tested in real conditions by the CEMS. It was notably used to assess the extent of the major floods that occurred this summer in Slovenia. The resolution of the images is currently limited to 15 meters, which makes it difficult to analyze certain images in city centers with narrow streets. Several avenues are being considered to improve this point as well as to improve this rapid mapping service more generally. “The images obtained are currently a little rough and we want to improve our system using digital terrain modeling (DTM) to achieve a resolution of the order of a few meters, adds Rémi Braun. We also want to introduce a hydraulic model to fill in certain gray areas in the images acquired by satellites. Finally, as the months go by, the AI ​​algorithm will become more and more efficient, because it will work from an increasingly large database. »

[ad_2]

Source link

Scroll to Top