A Multi-Model Benchmark for Archaeological Looting Pit Detection: From Fine-Tuned Detectors to Foundation Models

Archaeological looting pit detection from drone imagery poses a well-documented challenge in remote sensing archaeology, combining the need for high spatial precision with severe class imbalance and limited annotated training data. Our recent review (Melillos & Agapiou, Heritage 2026) identified domain-specific fine-tuning and curated benchmark datasets as the primary bottlenecks preventing multimodal AI from reaching operational status in archaeological workflows. This study presents the first empirical step beyond that review, moving from a theoretical framework to an operational multimodal detection pipeline. A field experiment was conducted over an open agricultural area, where drone imagery was collected across multiple flight plans under varying altitude and illumination conditions. In parallel, imagery from known looting case studies was analysed to ground the detection framework in real-world archaeological contexts. The proposed pipeline integrates and compares multiple AI paradigms for looting pit detection, including fine-tuned deep learning detectors, classical machine learning classifiers, and multimodal foundation models capable of operating in zero-shot settings without task-specific training. This multimodal approach allows for a systematic evaluation of how different model families respond to the visual characteristics of looting pits across varying acquisition conditions. Preliminary findings confirm the feasibility of automated detection and highlight the complementary strengths of fine-tuned and zero-shot models. This work contributes a first benchmark dataset and a reproducible multimodal evaluation framework for AI-assisted archaeological site monitoring.