Addressing the Accuracy Paradox in Archaeological Ceramic Surface Detection Using Low-Altitude, High-Resolution Remote Sensing D
Date Issued
December 2025
Author(s)
Advisor
Abstract
Archaeological surface surveys play a critical role in interpreting past human activities, yet conventional pedestrian approaches remain limited in both their spatial extent and the rate at which survey coverage can be achieved. The integration of high-resolution UAV imagery with machine learning techniques introduces new opportunities for the automated detection of surface sherds. However, a significant challenge remains, as ceramic sherds usually make up a minority of the pixels in an image. This class imbalance creates significant methodological difficulties that have so far hindered practical use.
This dissertation addresses the accuracy paradox in archaeological remote sensing: the phenomenon whereby machine learning classifiers achieve deceptively high overall accuracy (>95%) while systematically failing to detect minority-class archaeological features. This research develops and validates a comprehensive methodological framework explicitly designed for minority-class detection, integrating boosting algorithms (AdaBoost and XGBoost), threshold optimisation, and minority-aware evaluation metrics. Finally, based on the outcomes it proposes an index (Sensitivity Archaeological Detection -SAD) that can support future research.
The study employs data from three Mediterranean sites: Alampra (Cyprus) for controlled experimentation with 365 known ceramic placements, Kofinou (Cyprus) for archaeological validation with dense ceramic scatters, and Megara (Greece) for landscape-scale survey across four independent grids covering approximately 400 square meters. Both RGB and multispectral imagery were evaluated.
Results demonstrate consistent effectiveness across all phases. Phase I confirmed multispectral superiority under control conditions, achieving F1 = 0.76 (multispectral) versus F1 = 0.38 (RGB). In Phase II, boosting algorithms improved ceramic recall by approximately 60% relative to baseline methods (from 33% to 53% for RGB, and from 28% to 44% for multispectral). Critically, Phase III demonstrated that threshold optimisation—lowering decision boundaries from the default 0.5 to values between 0.11 and 0.34—substantially improves classifier performance: XGBoost and AdaBoost achieved F1-scores of 0.48–0.50 on independent validation data, representing a 75% cumulative improvement over Phase I baselines and exceeding the 0.35 operational viability threshold by 37–43%. Spatial validation at Megara confirmed that machine learning outputs correspond to field-verified ceramic concentrations, achieving 68% spatial recall with a 6.8× survey efficiency improvement.
This research establishes that the accuracy paradox in archaeological prospection can be overcome through the synergistic combination of boosting algorithms and calibrated decision boundaries. Multispectral imagery offers better spectral discrimination, as shown by its higher kappa (0.16 vs 0.07) and non-ceramic precision (92% vs 73%). However, the results also show that a threshold-optimized RGB workflow can still achieve operational performance (F1 ≈ 0.49) in contexts where multispectral acquisition is not feasible. The methodology thus offers complementary pathways: multispectral data for maximum precision, or optimized RGB for accessible, cost-effective deployment.
This dissertation addresses the accuracy paradox in archaeological remote sensing: the phenomenon whereby machine learning classifiers achieve deceptively high overall accuracy (>95%) while systematically failing to detect minority-class archaeological features. This research develops and validates a comprehensive methodological framework explicitly designed for minority-class detection, integrating boosting algorithms (AdaBoost and XGBoost), threshold optimisation, and minority-aware evaluation metrics. Finally, based on the outcomes it proposes an index (Sensitivity Archaeological Detection -SAD) that can support future research.
The study employs data from three Mediterranean sites: Alampra (Cyprus) for controlled experimentation with 365 known ceramic placements, Kofinou (Cyprus) for archaeological validation with dense ceramic scatters, and Megara (Greece) for landscape-scale survey across four independent grids covering approximately 400 square meters. Both RGB and multispectral imagery were evaluated.
Results demonstrate consistent effectiveness across all phases. Phase I confirmed multispectral superiority under control conditions, achieving F1 = 0.76 (multispectral) versus F1 = 0.38 (RGB). In Phase II, boosting algorithms improved ceramic recall by approximately 60% relative to baseline methods (from 33% to 53% for RGB, and from 28% to 44% for multispectral). Critically, Phase III demonstrated that threshold optimisation—lowering decision boundaries from the default 0.5 to values between 0.11 and 0.34—substantially improves classifier performance: XGBoost and AdaBoost achieved F1-scores of 0.48–0.50 on independent validation data, representing a 75% cumulative improvement over Phase I baselines and exceeding the 0.35 operational viability threshold by 37–43%. Spatial validation at Megara confirmed that machine learning outputs correspond to field-verified ceramic concentrations, achieving 68% spatial recall with a 6.8× survey efficiency improvement.
This research establishes that the accuracy paradox in archaeological prospection can be overcome through the synergistic combination of boosting algorithms and calibrated decision boundaries. Multispectral imagery offers better spectral discrimination, as shown by its higher kappa (0.16 vs 0.07) and non-ceramic precision (92% vs 73%). However, the results also show that a threshold-optimized RGB workflow can still achieve operational performance (F1 ≈ 0.49) in contexts where multispectral acquisition is not feasible. The methodology thus offers complementary pathways: multispectral data for maximum precision, or optimized RGB for accessible, cost-effective deployment.
File(s)![Thumbnail Image]()
Name
Argyro Argyrou PhD Theses.pdf
Size
6.35 MB
Format
Adobe PDF
Checksum (MD5)
de4b8c76c2581d4d360015d2da4e055a