Beyond Detection: Generating Risk Factors for Attempted Looting via Unsupervised Automated Algorithmic Detection

• In Event: Innovative Applications of Computational Methods

Abstract

Both manual and automated methods of detecting and documenting archaeological looting are well established. While the current literature has largely focused on establishing the most accurate and effective algorithmic approaches, less attention has been paid to what comes next. What should the resulting data be used for beyond detection and identification? The current study presents preliminary results on an ongoing interdisciplinary project attempting to address this question. Using eight years of very high resolution (VHR) satellite imagery (2015-2023) across 632 archaeological sites in the Nile Delta, this project compares a modified, existing unsupervised algorithmic approach (ALFEA) with a novel deep learning approach (SAMADA) using convolutional adversarial autoencoders. The best method (accurate, adaptable, scalable, sustainable) will incorporate a recurrent neural network (RNN) to link individual probable looting pits over time. The tool is being trained on multiple landscapes, site sizes, and contexts to maximize generalizability and will produce analytically useful spacetime data on archaeological looting. In this way, the resulting data will be useful for understanding the risk factors (social, environmental, economic, political) that predict looting activity.

Authors

• Michelle Fabiani, University of New Haven

• Shivanjali Khare, University of New Haven

• Thomas Powell, University of New Haven

• Jafar Vohra, University of New Haven

• Kalyna Yurchuk, University of New Haven