Potential of the Tree-based Pipeline Optimization Tool (TPOT) in detecting and mapping archaeological sites using remotely sensed images
By Serge Kiala,
Origins Centre MAEASaM Research Hub, Wits University, South Africa
Over the past few decades, there have been several challenges in studies that have attempted to map archaeological sites using conventional classifiers and remotely sensed images. For instance, studies have shown inconsistencies in the performance of compared algorithms or classifiers. Due to the site characteristics related to soil, or surrounding vegetation complexity, a classifier may be suitable for some sites and not for others. Generally, the identification of a data-independent algorithm remains a great challenge in applying remotely sensed images on mapping archaeological sites, because the determination of an optimal classification algorithm is often time-consuming and laborious, as it involves a comparison of several manually generated and complex trials. The process would also require a high skill set in machine learning that many non-specialist machine learning practitioners might not have. The Tree-based Pipeline Optimization Tool (TPOT) is an efficient method for selecting and tuning algorithms with limited human intervention.
The TPOT eases the process of choosing the machine learning and data pre-processing techniques to use for mapping archaeological sites in a given landscape. It can automatically map archaeological sites with higher accuracies than manually selected and parameter-tuned algorithms.
The TPOT is an open-source tool that has been rarely applied on remotely sensed images. By using the TPOT, it would likely be possible to find the best-performing classifier for every site. This is particularly important when mapping archaeological sites at regional scale where the mapping process is undertaken on different image tiles or when the mapping process are undertaken on remotely sensed images collected from different years. The use of one classifier may not be efficient in such case studies. The image below shows the output of a classified map of vitrified dung, a key archaeological marker for the presence of cattle enclosures, found in the Shashi-Limpopo Confluence Area (SLCA) of southern Africa. The work joins onto and expands on the remote sensing analyses conducted by Dr Merlo and Dr Thabeng. To read more visit: https://www.sciencedirect.com/science/article/pii/S0305440318300967
