A new breakthrough publication presents two algorithms for the large-scale automatic detection and instance segmentation of potential archaeological mounds on historical maps, resulting in the detection of nearly 6,000 mound features across an expansive area of 470,500 square kilometers. This remarkable achievement represents the most extensive application of such an approach to date and offers unprecedented opportunities for the reconstruction of ancient landscapes.
Over the past century, landscape modifications have been unprecedented, with the widespread implementation of mechanized agriculture, channel-based irrigation schemes, and urban expansion, among other factors. Historical maps provide a valuable glimpse into disappearing landscapes, depicting numerous historical and archaeological elements that no longer exist today.
The algorithms primarily focus on identifying and extracting mound features that exhibit a high probability of being archaeological settlements. Mounds are among the most frequently documented archaeological features found in the Survey of India historical map series, even though they might not have been recognized as such during the original survey.
Mound features with significant archaeological potential are often depicted through hachures or contour-equivalent form-lines. Consequently, the algorithms have been meticulously developed to detect each of these features. Our innovative approach addresses two common challenges faced in automated archaeological surveys: the sparse distribution of archaeological features for detection and the limited availability of training data.
These algorithms have been successfully applied to various types of maps within the historic 1-inch to 1-mile series, which has increased the complexity of detection. Leveraging synthetic data and employing a Curriculum Learning strategy, the algorithms have achieved a better understanding of mound features’ visual characteristics.
Additionally, a set of filters based on topographic setting, form, and size has been implemented to enhance the accuracy of the models. The resulting algorithms exhibit an impressive recall value of 66.96% and a precision of 96.25-88.51% (for high and low-density areas respectively) for hachure mounds, as well as a recall value of 70.59% and a precision of 100-92.31% (for high and low-density areas respectively) for form-line mounds. This breakthrough facilitated the detection of nearly 6,000 mound features across an expansive area of 470,500 square kilometers, marking the most extensive application of such an approach to date.
These remarkable algorithms represent a significant leap forward in the field of archaeological research and provide an invaluable tool for mapping and understanding ancient landscapes. The potential for uncovering previously unknown archaeological sites is immense, opening up new possibilities for scholars, historians, and archaeologists.
The study results from a collaboration between the Computational Archaeology Lab at the Catalan Institute of Classical Archaeology (ICAC-CERCA), The McDonald Institute for Archaeological Research of the University of Cambridge, and the Computer Vision Center of the Autonomous University of Barcelona (CVC-UAB).
Full reference
Berganzo-Besga, I., Orengo, H.A., Lumbreras, F. et al. Curriculum learning-based strategy for low-density archaeological mound detection from historical maps in India and Pakistan. Sci Rep 13, 11257 (2023). https://doi.org/10.1038/s41598-023-38190-x
Acknowledgements
The Mapping Archaeological Heritage in South Asia (MAHSA) project is funded by Arcadia, a charitable fund of Lisbet Rausing and Peter Baldwin. This research was also partially supported by Grant PID2021-128945NB-I00, awarded by MCIN/AEI/10.13039/501100011033, and by “ERDF A way of making Europe”. The authors acknowledge the support of the Generalitat de Catalunya CERCA Program to CVC and ICAC. Finally, the authors would like to thank Junaid Abdul Jabbar, Mou Sarmah, Ushni Dasgupta, Azadeh Vafadari, Kuili Suganya Chittiraibalan, Arnau Garcia-Molsosa and Adam Green.
