Abstract: Since the Middle Holocene, multiple ancient civilizations in the middle and lower reaches of the Yangtze River have constructed ancient dams for flood control and irrigation as a response to environmental changes. However, due to erosion, sedimentation and human activities, it is difficult for traditional archaeological methods to quickly identify and discover these ancient dams, which restricts the development of hydraulic archaeology. This paper proposes an efficient method for large-scale identification of ancient dams based on historical remote sensing images and deep learning technology, and evaluates the method by taking the ancient dams in Liangzhu Ancient City and its surrounding areas as research cases. This method collects aerial and satellite images from the 1940s to 1970s, labels 132 confirmed ancient dams in the study area, selects the YOLOv5 as the basic architecture, and optimizes the model by introducing the Generalized Intersection over Union (GIoU) loss function, coordinate attention mechanism and supplementary detection layer. The results show that the optimized model achieves a recall rate of 70% and a precision of 68% for ancient dams in the study area, which is significantly higher in efficiency and accuracy than traditional visual interpretation and field archaeological methods. This method provides an automatic tool for large-scale census of ancient water conservancy facilities, and is of great significance for revealing the spatio-temporal evolution characteristics of ancient water conservancy facilities, clarifying the human-land relationship model in which ancient people actively responded to environmental changes by constructing water conservancy projects under the background of climate change, and understanding the development process of ancient civilizations. 

Key words: ancient dams, historical remote sensing images, deep learning, Liangzhu Culture