Extraction of Linear Structures from LIDAR Images Using a Machine Learning Approach

Abstract

LiDAR (Light Detection And Ranging) technology makes it possible to generate highly accurate elevation models from the ground whatever the nature of the plant cover. LiDAR elevation models have proliferated during the past decade, delivering an unprecedented number of original archaeological finds in the forest. These include habitat, agricultural or funeral structures prior to the existence of forest cover, and also archaeological micro-structures directly linked to past forest economy.

Until recently, LiDAR acquisitions in France were limited to small areas. However, the recent and rapid supply of large-scale reference data by the National Geographic Institute provides large amounts of very high-resolution data about areas covering several thousand square kilometers that were previously little known from an archaeological point of view. Manual digitization of remains is a time-consuming activity and does not guarantee exhaustive recognition of features.

As part of the “SOLiDAR” project (a tribute to the federation of unions Solidarność) (http://citeres.univ-tours.fr/spip.php?article2133), we present a Machine Learning approach enabling reliable and flexible extraction and characterization of archaeological structures discovered in the LiDAR datasets. We have developed an open human-machine interface (HMI) that is accessible to the majority of archaeologists. This system, far from being a “black box”, can automatically process the remains but can also be used step by step, leaving the user to decide whether or not to validate the different processing parameters.