Semi-automated LiDAR Vegetation Classification for Mediterranean Archaeology: Designing a Pipeline Leveraging a Multi-Layer Stacked Ensemble Approach

Abstract. Dense Mediterranean vegetation often conceals archaeological features in LiDAR data, posing a significant challenge for archaeological analysis. This paper presents a novel machine learning pipeline for semi-automated vegetation classification in drone-based archaeological LiDAR point clouds, which were captured to survey the Mediterranean landscape of Sicily, Italy. Our approach integrates an extensive feature engineering stage with a multi-layer stacked ensemble classifier and a RandLA-Net deep learning model. The pipeline was trained on a semantically annotated drone-based LiDAR dataset from the site of Kamarina. It achieved high accuracy in distinguishing vegetation from ground points (0.99 overall accuracy, weighted macro F1 ≈ 0.93). To evaluate generalizability, we tested the model on a secondary site (Heloros) with different vegetation characteristics, obtaining an F1 of ~0.70. Qualitative inspection of results confirms that our model effectively removes vegetation while preserving archaeological structures. Our results demonstrate the potential of ensemble learning and 3D deep neural networks in archaeological remote sensing, enabling more efficient visualization and mapping of hidden archaeological features.