When we talk about environmental disasters, as could be the case of floods , different variables come into play: the difference in ground level, the amount of water received, as well as the thaw, just to name a few examples. It is precisely this diversity of data that makes it impossible to predict with any precision which areas will be flooded. And, above all, to what extent.
As has already been demonstrated on different occasions, the strength of AI tools lies especially in their ability to “learn” from the data sets that are sent to them, identifying trends from the links established between each of the variables.
In this way, Artificial Intelligence has already accelerated the process of detecting flooded buildings immediately after a large-scale flood , offering the possibility, to emergency personnel, to be able to direct their efforts in a much more efficient way.
Now, we have learned that a group of scientists from the University of Tohoku, in Japan, has created a machine learning model that uses photographs from the media in order to identify flooded buildings with greater precision within 24 hours. disaster, according to research published earlier this month in Remote Sensing magazine.
According to experts, this model would demonstrate how rapid media reports can be extremely useful in not only increasing but also speeding up the accuracy of damage mapping activities, offering the advantage of speeding up decisions. of emergency teams and optimize the response to these disasters .
Different algorithms for both machine learning and deep learning are designed for object classification through image analysis. But for machine learning and artificial intelligence itself to be truly effective, a series of data is necessary to train the model. And, in this specific case, these data would be those related to the flood.
Although they can be collected from previous events, they can generate quite a few problems because each event is different, and is also subject to the different local, specific and particular characteristics of the area that has been flooded. Hence, the information captured on the site through the communication media can have greater reliability .
It is clear that news crews are often the first to arrive at the scene of a disaster, broadcasting live images to viewers in their homes. And this is an aspect that the researchers have positively valued and recognized in their study, since it could be quite useful when used in artificial intelligence algorithms .
How does this model work?
The researchers applied their model to a city in Kurashiki, in Okayama prefecture, known as Mabi-cho, which in 2018 was affected by heavy rains in western Japan.
First, the researchers identified photos recognized and published by the press and geolocated them based on landmarks and other clues that appeared in the snapshots. They then used synthetic aperture radar images, provided by JAXA, to highlight flooded and non-flooded areas of unknown areas.
These types of images are especially interesting for classifying bodies of water, since microwaves tend to radiate differently depending on whether the surface is wet or dry.
Finally, the scientists used a support vector machine , which is one of the most common machine learning techniques, in order to classify buildings surrounded by floods, or located inside areas that had not been flooded.
And, according to their conclusions, the performance of the model reached an estimation precision of 80 percent .
