Etude du profil d'echelle des formes et de mesures d'energie de texture pour l'evaluation semi-automatique des degâts sur les bâtiments dans les images satellitaires de tres haute resolution.

摘要

Dozens of natural disasters occur each year throughout the world. They cause the death of thousands of people and cost billions of dollars in losses and reconstruction. Considerable resources are invested in the various phases of the emergency cycle like response and reconstruction. To avoid mismanagement of human and material resources during the response to a disaster, decisions must be made very quickly. To make appropriate choices, decision makers need up to date information on conditions on the ground. The images acquired by satellite are a possible source for this information. The new satellites equipped with optical sensors having spatial resolution finer than a meter per pixel provide details useful for determining the status of roads and damage to buildings. Unfortunately, visual analysis of these very large images is time consuming and fatigue may increase the rate of human error. In this thesis, we propose a semi-automatic method for the extraction of buildings and damage assessment using geometric, radiometric and texture features. The work includes a review of the literature in order to identify gaps in current approaches and possible solutions, the development of a methodology to solve this difficult problem, the testing of the proposed method on a portion of images of the city of Port-au-Prince, Haiti captured before and after the earthquake of January 12, 2010, assessment of results and their comparison with the literature. The proposed method requires processing of the image in a hierarchical tree shapes through the fast level set transform. Once the image is represented in this way, an algorithm for extracting meaningful forms is used to assign a representative form for each pixel. Geometrical descriptors such as area, perimeter and others from central moments are extracted from these forms. In addition, for damage assessment, energy measurements of texture are calculated on the forms before and after the event. Buildings are extracted using a supervised classifier based on a support vector machine (SVM). The damage is classified according to three degrees: little or no damage, damaged and destroyed. After experiment, the proposed method for the evaluation of damage exceeds those proposed in the recent literature both in rapidity and accuracy. The use of the scale profile and Laws textures are considerable innovations in the field of remote sensing brought by this thesis.