Moated site object detection using time series satellite imagery and an improved deep learning model in northeast Thailand

Abstract

Moated sites are crucial for revealing the formation of early civilizations and societies in Southeast Asia, and a significant amount of effort has been expended in investigating their distribution. This work is the first application of deep learning object detection methods to identify moated sites from time series satellite images. We presented multi-information fusion data (N-RGB) based on the fusion of multispectral and vegetation indices from Sentinel-2 time series imagery, generated a dataset of moated sites via the data augmentation method, and improved the YOLOv5s model by adding bidirectional feature pyramid network (BiFPN) structures for automatically identifying moated sites. The results indicate that the model trained with time series N-RGB data improves precision, recall, and mAP by more than 20.0% compared with single image data. The improved model was able to enhance the identification of small, moated sites and achieved 100% detection in a test of 100 moated sites. Ultimately, 629 targets were detected in northeast Thailand, with a false-negative rate of less than 3%, and 116 probable sites were identified. Among these, 6 probable sites were highly likely to be moated sites, as visually verified by high-resolution GEE imagery. In addition, among the targets automatically detected in other regions of continental Southeast Asia, the 5, 3, 2, 1, and 7 most probable sites were identified in Cambodia, Myanmar, Laos, Vietnam and other regions of Thailand, respectively. In summary, our approach enables the automatic detection of exposed and visible moated sites from satellite imagery, and could improve site discovery and documentation capabilities, opening new perspectives in larger geographic site units and even in civilization surveys.