Dams are critical hydraulic structures in arid environments to mitigate water shortages for sustainable regional water resource management and socioeconomic development. However, suitable sites for dams would change with global warming and sociodemographic development as the water supply and demand change spatiotemporally. This research develops a data-driven framework combining machine learning (Random Forest) and deep learning (YOLOv7-BiFormer) methods to explore the future optimal location selection of dams across large-scale regions based on multiple environmental and socio-demographic datasets. Focus on the Tarim River Basin, the "water tower" of Central Asia, where hundreds of hydraulic structures have been set up over the past decades and are considered to threaten the basin's hydrological and ecological security. 142 existing dams, including more than 100 unrecorded dams on the basin, are detected by applying the YOLOv7-BiFormer model to the basin through high-resolution remote sensing imagery (1.2 m). Our results show that cropland and runoff are key to affecting the site of dams, while elevation and climate are behind. The optimal sites of dams on the basin are mainly distributed in the Aksu and upper Yarkant rivers in the future under global warming. However, approximately ninety existing dams in the basin, especially in the Hotan and lower Yarkant rivers, would become useless and require removal by 2100. This research emphasizes the necessity for the management of dam sites in basins to foster the adaptation to social and climate change.