Extraction et évaluation des routes endommagées à partir d’images satellitaires par des algorithmes d’apprentissage automatique

Abstract

In recent years, the Earth has faced incidents and risks from major disasters of all types. These disasters cause enormous losses in terms of human life and infrastructure. The global community and governments have organized to prevent or cope with these dangers. One of the most effective means available to the international community and governments to curb this scourge is the use of remote sensing equipment which provide images almost instantly, in addition to providing a wide coverage of affected areas. When such a disaster occurs, photointerpreters produce maps showing the condition of operational and non operational infrastructures, allowing access to affected populations and makeshift hospitals, which are all crucial for rescue teams. This work shall contribute by assisting photointerpreters by automating their task in the image processing chain. The image processing chain consists in detecting essential infrastructures consisting of roads and to evaluate their level of damage. Firstly, for road detection modules, we are proposing jointly, a spectral angle algorithm (SA) based on gradient and multiresolution analyses. SA uses spectral information to distinguish roads from other forms of land cover, while MRA uses low-frequency linear (Gaussian) filters to reduce high frequency contents (noise) in a bid to simplify the original information for effective processing of the optical image. Secondly, and in the same vein, we will apply a combination of Bresenham algorithms (interpolation function for the linearity of pixels-roads) and the multiscale beamlets (multiscale analysis method) associated with the Frost filter to improve the detection of roads from synthetic aperture radar (SAR) images. SAR images have a noisy appearance (speckle) due to noise from backscattered waves. Frost filters considerably reduce speckle effects. Thirdly, we are proposing an improvement of road damage (blockage) detection, using the local binary pattern (LBP) and its variants through support vector machine (SVM) classifiers. These algorithms have successfully been applied on optic and radar image data. Our results were validated by an analysis showing the advantages and disadvantages of each of the approaches developed. This work was the subject of two papers and one journal article and submission (Annexe III).

Date	8 Sept 2026
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Michèle St-Jacques (Supervisor)