Thursday, 14 January 2016
Flash floods are one of the most severe disasters that occurs in the United States. For urban flash flood prediction, rainfall observation with high spatiotemporal resolution is required due to the complexity of urban hydrology systems and infrastructure. The existing rain observation network does not always have adequate spatiotemporal resolution for studying flash flooding. This study proposes to improve the spatiotemporal resolution of at-ground precipitation patterns by employing a novel technique based on the detection of rain intensity from images obtained from existing traffic cameras. New York City Department of Transportation (NYCDOT) cameras are used as a test bed for developing the initial technology. Traffic camera networks not only cover large parts of urban areas but also cover the critical hubs of ground transportation systems. The high density of these cameras in urban areas offers significant potential for improving the spatial detail of at-ground precipitation patterns. Furthermore, the NYCDOT cameras usually record images at a frequency of one image per two seconds, making the technique promising for measuring rainfall with a high temporal resolution. The proposed method estimates rainfall intensity based on high frequency fluctuations that the falling raindrops produce in the captured images. This signal manifests as additional image “noise” which can be extracted via computer vision techniques and correlated with ground truth rainfall intensity using traditional rainfall gauges. Based on the derived correlation, a model is built to estimate rainfall intensity based on the noise of a signal for each camera. The methodology presented in this paper is a fast, accurate, and inexpensive method for quantifying rainfall intensity using pre-existing camera networks.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner