Abstract Chronic flooding is an issue for low‐lying coastal communities globally, and it is expected to worsen with rising sea levels. Predicting when and where these floods occur can be difficult as they can be hyper‐local and ephemeral, depending on the flood drivers (e. g. , tides, rain). These factors make it difficult to measure the full spatial and temporal extent of chronic floods with in situ sensors. Here, we introduce a low‐cost (<400 USD), privacy‐preserving camera system that identifies flooding over block‐by‐block spatial extents at high frequencies (20 s–6 min). Our device—a Tiny Camera with machine learning (ML) (TinyCamML) —is a small, solar‐powered, microcontroller‐based camera that uses on‐device ML to classify images of roadways as containing a “flood” or “no flood. ” TinyCamMLs transmit only the classifications (a 1 or 0) to a website in real time, providing situation awareness during flood events over the entire image area while keeping data‐transmission costs low and preserving privacy. We demonstrate the TinyCamML's utility during both tidal and compound flood events in North Carolina, USA, which showed differences in flood spatial extents. During this deployment, the TinyCamML detected floods with an 81% accuracy, a 72% precision, and a 90% recall. The utility of the device extends beyond roadway flooding, as the onboard ML model can be easily retrained to capture other rare or ephemeral phenomena.
Farquhar et al. (Sun,) studied this question.