The Unseen Water Sink in Water-Stressed Regions: Automated Detection of Rooftop Cooling Towers for Urban Water Balances

Abstract Urban water management faces increasing challenges from rapid urbanization, climate change, and rising water demand, yet quantitative accounting of engineered evaporation sources remains incomplete. Rooftop cooling towers (CTs) represent a significant but often overlooked component of urban water fluxes, with limited spatial data available to support planning and policy. This study presents an integrated workflow combining YOLOv11 object detection, pretrained SAM2 segmentation, and a targeted human-in-the-loop (HITL) verification stage to identify and delineate CTs from high-resolution aerial imagery. The pipeline was trained on annotated imagery from Boston and tested at two locations near Phoenix, Arizona, achieving a high detection rate (recall of 96% and 82%), robust segmentation quality [intersection over union (IoU) of 76% and 67%], and requiring review of only a small fraction (3.5%) of total imagery to remove all false positives. The approach demonstrated label-free geographic generalization, maintaining performance across regions with markedly different climatic and architectural characteristics without fine-tuning. The resulting spatially explicit geometric measurements of fan radius should enable estimation of evaporative surface area for incorporation into urban water and hydrothermal models. This framework reduces manual annotation effort, lowers computational cost, and scales efficiently from local to regional applications, as demonstrated by its extension to Maricopa County. The method provides a transferable, operationally viable tool for integrating engineered evaporation sources into urban water budgets and supporting climate-resilient resource management.