Improving flood detection with large-scale dashboard camera data

Flooding poses a significant and growing challenge globally, threatening infrastructure, livelihoods, and public safety. However, current methods for detecting floods are limited in their spatiotemporal granularity and inequitable in their coverage. Here, we propose BayFlood, a method for fine-grained urban flood detection that identifies flooded street scenes in large-scale dashboard camera datasets using a vision-language model. We leverage the ability of modern vision-language models to identify floods even without large labeled datasets, which are typically unavailable. We comprehensively validate our approach using 1,440,184 images, showing that our model provides strong signal for floods across multiple cities and time periods and that our flood detections correlate with known external predictors of flood risk. We show our approach can be used to improve flood detection in New York City: our analysis detects floods in neighborhoods overlooked by current methods, identifies demographic biases in existing methods, and suggests locations for new flood sensors. This work underscores the potential of leveraging dense street-level imagery to significantly improve the understanding, detection, and management of flooding, and is more broadly applicable to detecting other objects and incidents from street scene data even when no labeled data is available.