Abdelaziz837/egyptian-archaeological-site-looting

Egyptian Archaeological Site Looting Detection (EASLD) Dataset

Dataset Description

This dataset contains high-resolution satellite imagery patches from Google Earth Pro historical imagery (2011–2017) covering multiple archaeological zones across Egypt. The primary objective is to facilitate the development and benchmarking of machine learning models to identify archaeological looting pits and protect cultural heritage.

Dataset Specifications

• Geographic Focus: Egypt (multiple archaeological areas of interest)
• Imagery Source: Google Earth Pro historical imagery (2011–2017)
• Data Format: JPG
• Resolution: Mixed spatial resolution
• Classes: 2 (Looted vs. Non-Looted)
• Dataset Size: 504 hand-annotated images

---

Academic Abstract

Archaeological looting poses an existential threat to cultural heritage preservation, particularly in historically dense regions like Egypt. While satellite remote sensing provides a scalable means of monitoring vulnerable sites, the development of automated detection pipelines is frequently hindered by a lack of open-access, labeled benchmarks. To address this gap, we present the Egyptian Archaeological Site Looting Detection (EASLD) dataset, a curated remote sensing dataset designed to train and benchmark machine learning models for looting detection. Based on methodology adapted from recent archaeological remote sensing studies, our dataset contains satellite imagery patches capturing both looted and pristine archaeological zones.

We provide baseline convolutional neural network (CNN) models that demonstrate the viability of computer vision in detecting spatial anomalies associated with illicit excavations, such as looting pits. By making both the dataset and codebase publicly available, we aim to lower the barrier to entry for machine learning researchers contributing to cultural heritage protection.

---

Baselines & Reference Code

The implementation pipeline is fully open-source and hosted on GitHub: Abdelaziz837/Satellite-Looting-Detection.

Baseline Architecture

• Model: ResNet-50 ensemble, pre-trained on the SSL4EO (Self-Supervised Learning for Earth Observation) dataset.
• Decision Threshold: Optimized to 0.4007 to minimize false negatives and prioritize active site protection.

Evaluation Metrics (504 Validation Samples via Out-of-Fold Cross-Validation)

Metric	Value	Detail
Accuracy	~83.5%	421 / 504 samples correctly classified
Precision (Looted)	~83.5%	Out of all flagged sites, percentage truly looted
Recall (Looted)	~84.4%	Percentage of actual looting detected
F1-Score (Looted)	~83.9%	Balanced harmonic mean of precision and recall

Confusion Matrix Breakdown

• True Negatives (Non-looted correctly identified): 204
• True Positives (Looted correctly identified): 217
• False Negatives (Looted sites missed): 40
• False Positives (Non-looted flagged as looted): 43

---

Citation & Licensing

If you use this dataset or code, please cite the underlying academic work:

• Based on research methodology adapted from: [https://www.cambridge.org/core/journals/antiquity/article/satellite-evidence-of-archaeological-site-looting-in-egypt-20022013/D23EA939FC4767D8BF50CAC6DE96D005]
• Dataset Citation: Abdelaziz837, Egyptian Archaeological Site Looting Detection (EASLD) Dataset, 2026.