DETECTING ANCIENT AGRICULTURAL TERRACES WITH DEEP LEARNING: A DATA FUSION APPROACH FROM THE BOZBURUN PENINSULA

Date

2025-7

Author

Peker, Emin Atabey

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

2421
views

0
downloads

Cite This

This study presents the first systematic application of deep learning to detect ancient agricultural terraces within a Turkish archaeological landscape, focusing on the Bozburun Peninsula in southwestern Anatolia. Historically part of the Rhodian Peraia during the Hellenistic period, the peninsula preserves extensive terrace systems shaped by centuries of agrarian activity. The research evaluates several UNet-based models—including early fusion, intermediate fusion, and late fusion architectures—alongside an RGB-only baseline. These models integrate spectral (RGB) and topographic (elevation, slope, aspect) inputs to identify terrace features using high-resolution aerial imagery (30 cm) and digital elevation models (DEMs). The study area spans 193 km² and encompasses diverse topographic conditions. Sixteen manually digitized sample areas (totaling 37.8 ha) were used to generate 256 image patches (512 ×512 pixels) for model training. Preprocessing techniques included RGB normalization (0–1 scaling) and circular normalization of aspect values via sine-cosine transformation. Among the tested models, the early fusion architecture incorporating topographic data achieved the best performance, with an average Intersection over Union (IoU) of 0.754 and an accuracy of 85.9%. A Monte Carlo evaluation with ten randomized initializations confirmed the model’s robustness across different fusion strategies. Spatial analysis revealed that 89.77% of detected terraces are located below 300 meters in elevation, predominantly on slopes of 10–20°, and are primarily north- to northwest-facing. The comparison with previous archaeological records showed a high degree of agreement in elevation and slope characteristics. Traditional approaches yielded higher recall (94.3% vs. 76.6%), whereas the AI model achieved slightly better precision (87.4% vs. 79.3%). Field verification at 27 locations further validated the model’s predictions. This research underscores the potential of deep learning for automated archaeological mapping in Mediterranean environments and provides a replicable framework for documenting vulnerable cultural landscapes such as the Bozburun Peninsula.

Subject Keywords

Bozburun Peninsula, Deep Learning, Agricultural Terraces, Remote Sensing, Digital Archaeology

URI

https://hdl.handle.net/11511/115149

Collections

Graduate School of Social Sciences, Thesis