Contribution à la segmentation d’images satellitaires à haute résolution avec une approche d’apprentissage machine. Application aux dunes de sable Barkhanes.

Abstract

The study of sand dunes is essential to understand and prevent the desertification phenomenon. Gathering data from the field is a labor intensive task, as deserts contain a large number of moving sand dunes. In this work, we contributed with a novel approach to automate their data collection, from High Resolution Satellite Images, using machine learning for image segmentation. We focused on Barchans as they were the fastest moving dunes. We faced many obstacles to segment these objects, due to the variability of their luminosity, size, orientation and non-rigid shape. To overcome these issues, we took advantage of image texture and used descriptors and statistical indicators, which provided good results. We improved them using a Machine Learning approach in which we combined a hierarchical cascade classifier for the detection of Barchans, with a new variation of Active Shape Model which we introduced for contour matching.. Moreover, we developed a new model for allometry. Finally, we generalized our solution, by taking into account colliding Barchans, which have erratic shapes, by introducing Transfer Learning, a variation of Deep Learning, to achieve the detection of Barchans dunes collisions. Our methods were tested on multiple high resolution satellite images. The experimental results were satisfying both in accuracy and robustness.