Objective comparison of relief visualization techniques with deep CNN for archaeology

Archaeology has been profoundly transformed by the advent of airborne laser scanning (ALS) technology (a.k.a airborne LiDAR). High-resolution and high-precision synoptic views of earth's topography are now available, even in densely forested environments, to identify and characterize landform patterns resulting from past human occupation. ALS-based archaeological prospection relies on digital terrain model (DTM) visualization techniques (VTs) that highlight subtle topographical changes perceived and interpreted by archaeologists. An increasing number of VTs have been developed, and they have been evaluated to date mainly based on subjective human perception. This study developed a new approach based on state-of-the-art computer-vision algorithms to benchmark VTs using objective metrics. Thirteen VTs were applied to a ALS-derived DTM, and a deep convolution neural network (deep CNN) was implemented and trained to automatically detect and segment archaeological structures from these images. Visual interpretation of the images showed that the most informative VT was e2MSTP, which combined a multiscale topographic analysis (MSTP) with a morphologically explicit image and a slope-invariant relief detrending technique. The deep CNN approach confirmed these results and provided objective performance metrics. This study indicates that the computer vision approach opens new perspectives in the objective selection of the most suitable VT for archaeological prospection.