Automatic detection of archaeological features from airborne laser scanning data using deep learning: Development and evaluation of the ADAF tool

The need to use machine learning (ML) in archaeology is constantly increasing due to the growing availability of large amounts of high-quality airborne laser scanning data (ALS, lidar). The Automatic Detection of Archaeological Features (ADAF) tool was developed to provide an easily accessible deep learning (DL) software for automating the detection of archaeological features from ALS data. The software has a user interface that requires minimal interaction and no prior knowledge of ML techniques, greatly improving accessibility for the archaeological community. The underlying ML models (MLMs) have been trained on an extensive archive of ALS datasets in Ireland, which have been labelled by experts with three types of archaeological features, namely enclosures, ringforts and barrows. The core components of the tool are the Relief Visualisation Toolbox (RVT) for processing the ALS data into ML-friendly visualisations and the Artificial Intelligence Toolbox for Earth Observation (AiTLAS), which provides access to the DL models, both of which are actively used in the field of aerial archaeology. Although developed for detecting selected archaeological monuments in Ireland, ADAF supports retraining with data from other regions and other monument types, enabling its use in archaeological contexts worldwide. The practical application of the software was tested on an active infrastructure development corridor in Ireland. The results showed a promising 84 % recall rate for known archaeological sites and outperformed human detection by discovering 116 potential new archaeological sites that had not been noticed during manual inspection. • ADAF is an easy-to-use deep learning tool for detection of archaeology from ALS data. • 10,000+ samples of enclosures, ringforts and barrows support robust model training. • HRNet and Faster R-CNN are used for semantic segmentation and object detection. • Real-world testing on 197 km 2 showed ADAF detected features missed in manual survey.