Upcoming Lecture On Machine Learning at AARG 2024
We are pleased to announce that Alfie Leek will present a paper at the 41st annual Aerial Archaeology Research Group conference on the 13th of September. The paper titled:
An assessment of the efficacy of automated feature detection in Archaeology using National LiDAR Programme data for the 12th Century monastic landscape at Ravenstonedale, Cumbria.
Abstract:
The monastery and advowson were donated to the Gilbertine Order of Sempringham sometime during the reign of Henry II, 1145–1189. While we know that the Gilbertines lost control of the monastery at some point during the dissolution of 1536–1541, there is little understanding of the monastery and its role in the broader landscape. This is due primarily to poor preservation and the fact that after the Reformation, most of the documents regarding the monastery were kept in the Tower of St Mary’s Abbey, which was subsequently blown up by the Parliamentarian army during the siege of York in 1664. From 1928 to 1929, only a limited archaeological excavation was done on conventional monastic buildings, with a recent test pitting survey being conducted in the town. This project follows on from the aerial survey I delivered for Archaeological Research Services on behalf of Historic England and the Yorkshire Dales National Park Authority in 2022–2023.
Automated feature detection, a promising field of investigation within remote sensing, has the potential to enhance research and cultural heritage management significantly. This paper aims to assess the efficacy of both pixel and object-based machine learning algorithms in detecting two feature types associated with Ravenstonedale Monastery, Cumbria: warrens and water meadows. Despite the varied geology of the study area, which is known to pose a challenge for automated approaches, we could accurately predict a significant portion of the archaeology. Whilst visually distinct, the warrens' small scale and geology variation meant that LiDAR pixel-based approaches struggled due to homogenous colouration between features and geology. Conversely, pixel-based approaches successfully detected the water meadow’s ditched features due to the clarity of the leat and channels as well as the relatively level geology. Object-based approaches could detect between 90.5% and 28.6% of the warrens, depending on the degree of filtering to reduce overprediction.
The results of these successful approaches will be discussed and compared to traditional aerial survey techniques, offering a hopeful outlook on the future of archaeological research. This discussion will be split into two parts, first focusing on single vs multi-variable automation, presenting the benefits of both, and the reasons why multi-variable is more accurate. The subsequent discussion will explore whether automated feature detection has been over-promised or is too early. Both are true, as machine learning is still in its infancy and approaches to automated feature detection should be viewed in Gartner’s hype cycle framework. Ultimately, automated feature detection is useful for both commercial and research archaeology when deployed correctly, but it should not be viewed as omnipotent. Further development of processes and teaching resources would only benefit the field of aerial archaeology going forward.Be clear, be confident and don’t overthink it. The beauty of your story is that it’s going to continue to evolve and your site can evolve with it. Your goal should be to make it feel right for right now. Later will take care of itself. It always does.