Barchan Sand Dunes Collisions Detection in High Resolution Satellite Images based on Image Clustering and Transfer Learning

Abstract

Desertification is a core concern for populations living in arid and semi-arid areas. Specifically, barchans dunes which are the fastest moving sand dunes put constant pressure over human settlements and infrastructure. Remote sensing was used to analyze sand dunes around Tarfaya city located in the south of Morocco in the Sahara Desert. In this area, dunes form long corridors made of thousands crescent shaped dunes moving simultaneously, thus, making data gathering in the field very difficult. A computer vision approach based on machine learning was proposed to automate the detection of barchans sand dunes and monitor their complex interactions. An IKONOS high resolution satellite image was used with the combination of a clustering algorithm for image segmentation of the dunes corridor, and a Transfer Learning model which was trained to detect three classes of objects: Barchan dunes, bare fields and a new introduced class consisting of dunes collisions. Indeed, collisions were very difficult to model using classical digital image processing methods due to the large variability of their shapes. The model was trained on 1000 image patches which were annotated then augmented to generate a larger dataset. The obtained detection results showed an accuracy of 84,01%. The interest of this research was to provide with a relatively affordable approach for tracking sand dunes locations in order to better understand their dynamics.