Archaeoscape: Bringing Aerial Laser Scanning Archaeology to the Deep Learning Era

TL;DR

Archaeoscape introduces the largest open-access ALS archaeology dataset to date, spanning 888 km^2 in Cambodia with 31,411 annotated features to enable deep-learning semantic segmentation under dense vegetation. The paper benchmarks a broad set of contemporary segmentation models, including CNNs and vision transformers, illustrating that elevation-derived channels are often more informative than RGB in canopy-rich environments and that larger context improves performance, though results remain below CV benchmarks, underscoring domain-specific modeling challenges. It also details a careful misuse-mitigated data policy (parceled, obfuscated geolocation, credentialed access) and provides extensive acquisition, preprocessing, and annotation pipelines to support reproducibility and future research. Overall, Archaeoscape aims to bridge archaeology and computer vision, providing a rigorous benchmark and a call for novel algorithms tailored to ALS data in complex cultural landscapes.

Abstract

Airborne Laser Scanning (ALS) technology has transformed modern archaeology by unveiling hidden landscapes beneath dense vegetation. However, the lack of expert-annotated, open-access resources has hindered the analysis of ALS data using advanced deep learning techniques. We address this limitation with Archaeoscape (available at https://archaeoscape.ai/data/2024/), a novel large-scale archaeological ALS dataset spanning 888 km$^2$ in Cambodia with 31,141 annotated archaeological features from the Angkorian period. Archaeoscape is over four times larger than comparable datasets, and the first ALS archaeology resource with open-access data, annotations, and models. We benchmark several recent segmentation models to demonstrate the benefits of modern vision techniques for this problem and highlight the unique challenges of discovering subtle human-made structures under dense jungle canopies. By making Archaeoscape available in open access, we hope to bridge the gap between traditional archaeology and modern computer vision methods.

Figures (7)

• Figure 1: Archaeoscape. Our proposed dataset contains 888 km$\textsuperscript{2}$ of aerial laser scans taken in Cambodia. The 3D point cloud LiDAR data (left) was processed to obtain a digital terrain model (middle). Archaeologists have drawn and field-verified 31,411 individual polygons by delineating anthropogenic features (right).
• Figure 2: Archaeoscape overview. We show the vectorial annotations overlaid onto the relative elevation maps for each parcel, and their assignment to the training, validation or test splits. The position and orientation of the parcels is arbitrary. The geometry of the annotations has been simplified to reduce the file size of the paper. Best viewed on a computer screen.
• Figure 3: Archaeoscape classes. We illustrate the three main classes with in-situ images (top row), top-view hillshaded elevation maps (middle row), and our annotations (bottom row). In many cases, the sought features are difficult to detect visually by in-situ observation but are more apparent on elevation maps.
• Figure 4: Qualitative performance. We provide examples of input elevation maps (\ref{['fig:quali:a']}) and their corresponding annotations (\ref{['fig:quali:b']}), as well as the prediction of a standard U-Net (\ref{['fig:quali:c']}) and our best model (\ref{['fig:quali:d']})---improvements in green. The red squares represent the size of the input images: 224 pixels, or 112m.
• Figure 5: Channel ablation. We represent the orthophotography (\ref{['fig:quali_ablation:a']}), normalized terrain model (\ref{['fig:quali_ablation:b']}), and annotations (\ref{['fig:quali_ablation:c']}). We also provide the prediction of a PVTv2 model operating on RGB photos (\ref{['fig:quali_ablation:d']}), and a model processing both RGB and elevation data (\ref{['fig:quali_ablation:e']}). The model using only radiometric information performs worse overall, and in particular, fails to identify any structures under the heavily forested area at the top left corner.