Revealing Lichen Covered Rock Carvings At Stonehenge

Another paper on the fascinating and innovative discovery methods of identifying rock art at Stonehenge that is hidden under the lichen has been released;

Gavin Leong , Matthew Brolly , David J. Nash , Novel lichen simulation and laser scan modelling to reveal lichen-covered carvings at Stonehenge, Results in Engineering (2025), doi: https://doi.org/10.1016/j.rineng.2025.106377 

This follows the pre-print released earlier this year:

Leong, Gavin and Brolly, Matthew and Anderson-Whymark, Hugo and Nash, David and Bedford, Jon, Novel Approaches for Enhanced Visualisation and Recognition of Rock Carvings at Stonehenge. Available at SSRN: https://ssrn.com/abstract=5126093 or http://dx.doi.org/10.2139/ssrn.5126093

and Gavin Leong's thesis:

Leong, G. (2024) Revealing lichen-covered rock art at Stonehenge: where terahertz imaging and photogrammetry meet machine learning, lichen simulation, and image analysis algorithms. PhD thesis, University of Brighton. Available at: https://research.brighton.ac.uk/files/49055416/PhD_Thesis_Gavin_Leong.pdf

Gavin Leong's PhD thesis, titled "Revealing Lichen-Covered Rock Art at Stonehenge: Where Terahertz Imaging and Photogrammetry Meet Machine Learning, Lichen Simulation, and Image Enhancement" and submitted in May 2024, serves as the foundational bedrock for the subsequent research outputs, including the February 2025 preprint on enhanced visualisation and recognition of exposed rock carvings and the July 2025 journal pre-proof on lichen simulation for detecting obscured carvings, both of which he led as the primary author. Drawing from extensive fieldwork at Stonehenge and innovative interdisciplinary approaches, Leong's thesis meticulously developed key methodologies such as the difference of Gaussians (DoG) and pseudo-depth mapping (PDM) techniques for visualising faint carvings on non-lichenised surfaces, achieving the discovery of new axe-head motifs on Stone 53, while pioneering the application of MeshNet—a 3D shape classification neural network—for semi-automated carving recognition with 90.7% accuracy on photogrammetry-derived meshes; these elements directly underpin the visualisation-focused preprint, providing the baseline tools and empirical discoveries that were refined and cited therein. Furthermore, the thesis introduced the groundbreaking Ramalina siliquosa diffusion-limited aggregation (RDLA) simulation, informed by laser-scanned lichen thickness data from Stone 30, to digitally replicate lichen occlusion on carving meshes, enabling denoising strategies that reduced visual noise by 70.7% and facilitating MeshNet retraining for 73.3% accuracy on simulated lichen-covered surfaces—innovations that form the core of the later journal article, extending the thesis's non-invasive ethos to address the 23% lichen-covered areas at Stonehenge without physical removal. Leong's comprehensive exploration of terahertz time-domain spectroscopy (THz-TDS) in laboratory and fieldwork settings, including partial least squares regression for substrate recovery beneath lichen layers up to 35 mm thick and at water contents below 18%, not only validated subsurface imaging as a complementary tool but also highlighted ethical conservation challenges, laying the groundwork for broader applications in global rock art sites and related fields like forest canopy modelling. Through rigorous experimentation, detailed appendices on code implementations, and a synthesis of machine learning with heritage science, Leong's doctoral work not only anticipated the challenges of lichen obscuration but also provided the theoretical, methodological, and empirical foundations that enabled the two papers to advance the field, deserving full credit for his pioneering contributions that bridge archaeology, remote sensing, and computational simulation in a manner that promises lasting impact on cultural heritage preservation.

The February 2025 preprint, titled "Novel approaches for enhanced visualisation and recognition of rock carvings at Stonehenge," introduced innovative visualisation techniques such as difference of Gaussians (DoG) and pseudo-depth mapping (PDM) to enhance the detection of faded carvings on exposed surfaces, leading to the discovery of four new carvings, ten potential carving areas, and nine alternative interpretations on Stone 53, while also demonstrating the efficacy of the MeshNet neural network for semi-automated carving recognition with 90.7% accuracy on non-lichen-covered meshes derived from photogrammetric data; it further suggested future extensions to lichen-obscured regions.

The later paper from July 2025, titled "Novel lichen simulation and laser scan modelling to reveal lichen-covered carvings at Stonehenge," explicitly builds upon the foundational preprint from February 2025 by addressing a key limitation identified in the earlier work—namely, the inability to analyse approximately 23% of Stonehenge's stone surfaces due to coverage by the fruticose lichen Ramalina siliquosa, which potentially obscures additional Early Bronze Age axe-head and dagger carvings. In contrast, the later paper extends this by developing a novel species-specific lichen simulation called Ramalina siliquosa diffusion-limited aggregation (RDLA), which modifies base diffusion-limited aggregation (DLA) with cone-based attachment (CBA) modes, logistic functions for branching behaviour, and efficiency improvements informed by new laser scan data from Stone 30 to capture real-world lichen thickness distributions (up to 37.5 mm, far exceeding typical carving depths of less than 1 mm). This simulation is applied to the photogrammetry-derived carving meshes from the earlier study (referred to as seed meshes), creating a database of digitally lichen-covered surfaces; subsequent denoising techniques, including cloth simulation sheathing and distance mapping, reduce lichen-induced visual noise by 70.7% as quantified by probability of superiority metrics, enabling clearer visualisation of underlying carvings via depth maps without physical lichen removal. A standout new result is the retraining and testing of MeshNet on these simulated lichen-obscured meshes, achieving 73.3% accuracy in distinguishing carvings from non-carvings (with high recall at 86.0% but lower precision at 68.5%), demonstrating that laser scanning combined with denoising and MeshNet remains viable for non-invasive detection even under shrubby lichen cover, albeit with reduced performance compared to the 90.7% baseline on clean data from the preprint. Additionally, the paper provides empirical data on lichen thicknesses from multiple species at Stonehenge, highlighting R. siliquosa's unique obscuring potential, and discusses broader applications, such as adapting RDLA for other fruticose lichens, subsurface imaging alternatives like XRF or terahertz, and extensions to global rock art sites or even unrelated fields like forest canopy modelling and virtual environment ageing effects, thereby advancing ethical conservation practices and opening avenues for further discoveries at Stonehenge and beyond.