TL;DR
Researchers have developed a hybrid machine learning pipeline that combines classical ensemble methods with deep learning to strip away dense Mediterranean vegetation from drone-based LiDAR data. By integrating XGBoost, CatBoost, and a Multi-Layer Perceptron with the RandLA-Net architecture, the system achieves a 0.99 accuracy in revealing hidden archaeological structures that standard filtering tools often miss.
What Problem Does This Paper Address?
In archaeological remote sensing, LiDAR (Light Detection and Ranging) is the gold standard for "seeing through" forests to find ruins. However, the "Mediterranean scrub"—a dense tangle of small trees, thorny shrubs, and bushes common in places like Sicily—presents a unique nightmare.
Standard "off-the-shelf" vegetation filters often struggle with this landscape. They frequently misclassify dense shrubs as ground or, worse, label actual archaeological walls and structures as vegetation, effectively deleting the history researchers are trying to find. This paper addresses the need for a tailored, semi-automated classification pipeline that can distinguish between complex Mediterranean flora and human-made remains with high precision.
Methods
The authors propose a multi-stage pipeline that blends "hand-crafted" feature engineering with state-of-the-art neural networks:
- Preprocessing & Feature Engineering: The raw point cloud is voxelized (downsampled to 15cm cubes) to manage data volume. For every point, the team calculates a "feature vector" including geometric descriptors (like planarity, sphericity, and verticality) and radiometric data (RGB values and LiDAR intensity).
- The Stacked Ensemble (Layer 1): Instead of picking one model, they use three: XGBoost, CatBoost, and a Multi-Layer Perceptron (MLP). Their individual probability outputs are fed into a Logistic Regression meta-classifier. This "stacking" approach creates a more robust initial guess.
- Deep Learning Integration (Layer 2): The real innovation is feeding those ensemble probabilities plus the original coordinates and features into RandLA-Net (Randomly Subsampled Local Aggregation Network). This allows the model to use the "wisdom" of the ensemble while also considering the broader spatial context of the point cloud.
- Interpolation: Finally, an Inverse Distance Weighted (IDW) method projects the classifications from the voxelized points back onto the original high-resolution dataset.
Datasets
The study utilized two primary sites in Sicily, scanned using a RIEGL VUX-1UAV22 sensor mounted on a heavy-lift drone:
- Kamarina: A 15-hectare area of an ancient Greek settlement. This served as the primary training and validation set, containing ~116 million points (15% vegetation, 85% ground/structures).
- Heloros (Marianelli area): A smaller, 0.03 km² site used strictly for testing generalization. This site had a much higher vegetation density (~65%), providing a challenging "out-of-distribution" test for the model.
Key Insights and Ideas
- The Hybrid Advantage: The authors found that while the ensemble alone was decent, adding RandLA-Net significantly boosted the F1-score for vegetation from 0.51 to 0.86. This suggests that "spatial awareness" (the DL component) is critical for identifying irregular vegetation patterns.
- Feature Synergy: Combining spectral data (how green a point is) with geometric data (how "spherical" or "scattered" a cluster of points is) proved highly effective for the "Mediterranean scrub" which lacks the clear vertical structure of a typical forest.
- The "Ground Truth" Paradox: A fascinating insight from the Heloros test was that the model's quantitative performance (F1 ~0.70) looked worse than its qualitative performance. Visual inspection showed the model was actually finding vegetation that human annotators had missed or mislabeled, highlighting the difficulty of creating perfect training data in dense terrain.
Key Results
- Kamarina (Holdout Set): The full pipeline achieved an overall accuracy of 0.99 and a weighted macro F1-score of 0.93.
- Vegetation Precision: In the training site, the model reached a precision of 0.79 for vegetation, meaning it rarely "deleted" archaeological structures by mistake.
- Generalization: When applied to the new Heloros site without any retraining, the model maintained an accuracy of 0.70. While lower than the training site, visual results confirmed it effectively cleared main vegetation clusters while preserving the underlying terrain.
Limitations and Future Work
The authors identify a few hurdles for scaling this approach:
- Labeling Quality: Manual labeling is prone to error in transition zones (where bush meets ground). Better training data is needed to reduce the "noise" the model learns.
- Computational Cost: Calculating neighborhood features for millions of points is resource-intensive, creating a barrier for researchers without high-end hardware.
- Sensor Fusion: The team plans to incorporate Long-Wave Infrared (LWIR) imagery. Since vegetation and stone structures have different thermal signatures, this could provide the "missing link" for even more accurate classification.
- Foundation Models: They are looking into fine-tuning large-scale pre-trained LiDAR models (like IGNF’s RandLA-Net) to see if they offer better out-of-the-box generalization for diverse landscapes.
Summary updated