AUTOMATED DETECTION OF STONE-WALLED RUINS USING BASED ON SUPPORT VECTOR MACHINE AND HISTOGRAM OF ORIENTED GRADIENTS

Amandine ROBIN, Karim SADR

Abstract


Aerial or satellite imagery allow archaeological surveys of large areas for a fraction of the time and cost of ground surveys. Nevertheless, the task of examining reams of air photographs or zooming into details on Google Earth is also a time consuming exercise. Therefore a desirable objective is to find a way of automating the detection of archaeological sites on remotely sensed imagery.

This new method proposes an autonomous approach to detect ruins, based on Histograms of Oriented Gradients for feature extraction and on a Support Vector Machine in order to classify the extracted features into a ruin vs non-ruin class. The support vector machine uses a training set to learn to distinguish the ruins from the rest, and is then applied to a wide area without any a priori knowledge to detect the ruins.
The approach is validated over the Suikerbostrand area in South Africa, to identify and classify pre-colonial stone-walled structures in an 8000 km2 study area. Over 7000 structures have been identified by a team of research assistants and are used as ground truth. The main challenges specific to this context are that the structures we seek to detect are very subtle and made from locally available material, shapes are diverse and tend to be occluded by other features such as vegetation. Thus, ruins are difficult to differentiate from natural features.

The performances of the method are analyzed depending on the set used for training, and the use of satellite images (LANDSAT, from Google Earth) vs LIDAR images is discussed. In both cases, the results demonstrate the relevance of this approach with a very good level of accuracy (more than 85%) and a good control of the false detections.