Automatic Detection of Reflections at Ancient Walls in Simple-Offset GPR Images by Using Cascade Classifiers

Simple-offset GPR reflection methodology allows obtaining very precise information in archaeological/historical sites. However, as large amounts of data are usually acquired, their processing, analysis and interpretation can be extremely time-consuming.

In this work, we present three algorithms for the automatic detection of reflections at ancient walls in SO-GPR images, based on cascade classifiers and well-known image feature descriptors: Haar, HOG and LBP. These algorithms were implemented using supervised learning, and experimental data from previous works. The best performances corresponded to the descriptor Haar. With only two cascade stages, remarkably accurate results were attained despite the complex characteristics of the signals of the walls. Almost all of them were detected near their actual positions, and only a few false positive predictions were obtained, mostly without any continuity across the profiles.

The main advantage of these methodologies is that once an accurate and reliable algorithm is implemented using data from an appropriate sector, it can be applied in all the zones of the site with similar characteristics, or even in other site of the same type. Thereby, a precise representation of the target structures is rapidly obtained, and the qualified interpreter only has to examine some parts of particular profiles.