Mapping Hidden Heritage: Self-supervised Pre-training for Archaeological Stone Wall Mapping in Historic Landscapes Using High-Resolution DEM Derivatives

TL;DR

This work tackles the challenge of mapping undocumented dry-stone walls in complex, vegetated landscapes where labeled data are scarce. It introduces DINO-CV, a cross-view self-supervised pre-training framework that learns view-invariant representations from high-resolution DEM derivatives (MHS and VAT) to improve wall segmentation when fine-tuning data are limited. Through experiments on the Budj Bim Cultural Landscape, DINO-CV outperforms supervised and prior self-supervised baselines across ConvNet and ViT backbones, achieving up to 68.6% mIoU with full data and maintaining strong performance with as little as 10% labeled data (e.g., 63.8% mIoU). The approach offers a scalable, data-efficient solution for large-scale cultural heritage mapping and environmental monitoring in inaccessible or sensitive regions.

Abstract

Historic dry-stone walls hold significant cultural and environmental importance, serving as historical markers and contributing to ecosystem preservation and wildfire management during dry seasons in Australia. However, many of these stone structures in remote or vegetated landscapes remain undocumented due to limited accessibility and the high cost of manual mapping. Deep learning-based segmentation offers a scalable approach for automated mapping of such features, but challenges remain: the visual occlusion of low-lying walls by dense vegetation and the scarcity of labeled training data. This study presents DINO-CV, a self-supervised cross-view pre-training framework based on knowledge distillation, designed for accurate mapping of dry-stone walls using high-resolution Digital Elevation Models (DEMs) derived from airborne LiDAR. By learning invariant structural representations across multiple DEM-derived views, specifically Multi-directional Hillshade (MHS) and Visualization for Archaeological Topography (VAT), DINO-CV addresses both occlusion and data scarcity challenges. Applied to the Budj Bim Cultural Landscape (Victoria, Australia), a UNESCO World Heritage site, the approach achieves a mean Intersection over Union (mIoU) of 68.6% on test areas and maintains 63.8% mIoU when fine-tuned with only 10% labeled data. These results demonstrate the potential of self-supervised learning on high-resolution DEM derivatives for large-scale, automated mapping of cultural heritage features in complex and vegetated environments. Beyond archaeology, this approach offers a scalable solution for environmental monitoring and heritage preservation across inaccessible or environmentally sensitive regions.

Figures (9)

• Figure 1: Overview of the airborne LiDAR dataset covering 301 km2 of the Budj Bim Cultural Landscape in southeastern Australia. The survey captured over 33 billion points using multi-return laser scanning, providing the high-resolution topographic data required for generating digital elevation models (DEMs) and derivative visualizations used in this study.
• Figure 2: The aerial image (Aerial, Row 1), Multi-directional Hillshade (MHS, Row 2), and Visualization for Archaeological Topography (VAT, Row 3) of dry-stone walls in the Budj Bim landscape. From left to right, columns show: (1–2) multiple parallel thin walls and an isolated thin wall; (3–4) thick, zigzag walls and a straight wall partially occluded by vegetation; (5–6) roadside straight walls with gate openings, and a heavily occluded wall segment.
• Figure 3: (a) The DINO-CV framework for self-supervised pre-training. At each training iteration $k$, the student network $g_{\theta_s}$ receives a randomly sampled input from the collection of DEM-derived views (i.e., MHS and VAT), while the teacher network $g_{\theta_t}$ receives a fixed view (MHS). (b) The Pseudo Siamese Network used for supervised fine-tuning. This architecture comprises two teacher networks, each pre-trained with DINO-CV on a distinct input view (MHS and VAT), enabling the extraction of complementary features for stone wall mapping.
• Figure 4: Spatial partitioning of the BudjBimArea dataset. Orange lines indicate annotated European historic dry-stone walls near Tae Rak (Lake Condah), Victoria, Australia. The number of data samples per area is as follows: Area 1 - 186, Area 2 - 1128, Area 3 - 1138, Area 4 - 1233, Area 5 - 1604, and Area 6 - 1813.
• Figure 5: Test results (mIoU%) of each vision backbone on BudjBimArea dataset (area 1-6).