Harnessing Computer Vision for Advanced Flood Forecasting in Urban Environments

Urban pluvial flooding, caused when local precipitation intensity outpaces the capacity of natural and engineered drainage systems, is among the costliest, most dangerous, and most prevalent forms of natural disaster. While flood inundation extent maps form the basis for federal management of coastal and riverine flood risk management, there is no existing technology that allows for real-time spatially continuous monitoring of urban pluvial flooding, which critically limits equitable and effective planning, response, and mitigation. The proposed research enables a low-power distributed sensor network that synchronizes storm sewer stage monitoring with a camera-based flood mapping platform that will provide automatic monitoring and alerts of urban surface inundation and stormwater backups, and autonomously generate spatially complete maps of peak flooding extent for comprehensive and equitable pluvial risk characterization and mitigation. To realize this vision, applied research will be conducted to leverage edge-AI (Artificial Intelligence) to identify flooding in multispectral images representing diverse urban and peri-urban contexts. Efficient methods for georeferencing images using photogrammetric models of the neighborhoods where cameras are installed will be developed and evaluated. By synchronizing cameras with LoRaWAN-networked USGS storm sewer stage sensors, we will demonstrate how edge-AI enables spatially distributed inundation measurements from cameras with low energy and communication costs and without collecting raw imagery that could contribute to unnecessary surveillance in host communities. Furthermore, we will demonstrate how networked sensor platforms can get smarter over time.