Disaster damages detection with convolutional joined auto-encoders

Today, with the rise of the new space sector, the data flows generated by satellites are becoming colossal. To meet the challenge of processing and analyzing them, the development of tools based on artificial intelligence is a major challenge. Thanks to the progress made in these two areas, humanity reaches new ways of managing natural disasters. The very first need after a disaster is to be able to geolocate precisely the damages to focus search, rescue and logistic decisions in order to save lives. The geolocation of this damage is based on the comparison of an image preceding the disaster and another one after the disaster. While the principle of comparison is simple, its concrete implementation requires the resolution of a number of complex problems. To solve them, Jérémie Sublime and Ekatarina Kalinicheva have developed a disaster damages detection model which they consider to be particularly effective because of the very specific and unique structure. The objective of this thesis is to confirm the efficiency of this model - and possibly improve it - by testing it on three datasets chosen from the Open Data Program of Maxar. After having underlined the importance of using the right metrics adapted to a specific situation, a first test with the closest model as possible had been made. It did not show equivalent results. Further analysis on the impact of the parameters had been made. None of them had shown drastic improvements. However, this work proposes a way to speed up the computational time using random initializations and smaller samples of batches as well as interesting insights that deserves further analysis.