Segmentation of crescent sand dunes in high resolution satellite images using a support vector machine for allometry

Abstract

The study of sand dunes movement is essential to understand and prevent the desertification phenomenon, and collecting data from the field is a labor intensive task, as deserts contain usually a large number of sand dunes. We propose to use computer vision and machine learning algorithms, combined with remote sensing and specifically high resolution satellite images for collecting data about the position and characteristics of moving sand dunes. We focused on the fastest moving sand dunes called barchans, which are threatening the settlements in the region of Laayoune, Morocco. We developed a process with three stages: In the first stage, we used an image processing approach with cascading Haar features for the detection of dunes location. In the second stage, we used a support vector machine for the segmentation of contours, and in the third stage we used an algorithm to measure the allometric features of barchans dunes. We explored the collected data, and found relevant correlations between dunes length, and width, and horns sizes, which could be used as key indicators for dunes growth and progression. This study is therefore of high interest for urban planners and geologists who study sand dunes and require technical methods, based on machine learning and computer vision to allow them to collect large amount of data from satellite images to understand sand dunes progression and counter desertification problems. The use of cascading Haar feature provided a good accuracy, and the use of Support Vector Machines, along with the high resolution satellite images provided a good precision for the segmentation of barchan dunes contours, allowing the collection of morphological features which provide significant information on barchans sand dunes dynamics.