Fast and Flexible Composition of Animated Flood Surfaces into Street-Level Photographs

Abstract:

Flooding is a frequent danger to communities located near coasts and rivers. Communicating the threats of potential floods to these at-risk populations is critical to ensuring their safety. However, traditional approaches that warn residents of forecasted floods are often ignored or are not fully understood, so new methods are needed to increase the effectiveness of flood communications. Here we propose an efficient and scientifically-grounded approach to generate realistic images and animations of a flood at any height composited with photographs taken at street level. With vehicular LIDAR point cloud and RGB photo data, we employ a convolutional neural network to perform dense depth completion on each image. We then use Blender to render a water surface at the appropriate height and orientation relative to the camera, which we superimpose onto the original photo. Using data from the Kitti Vision Benchmark Suite as a demonstration, we show the resulting visualizations allow for solid objects, such as walls, cars, and street signs, to occlude the water behind them, while also accounting for ground slope. For each scene, we generate over 180 such image compositions with increasing water heights and moving waves, which are then compiled into an animated video. Using a 4-core CPU and an NVIDIA Tesla K80 GPU, the neural network takes about seven seconds per scene to complete the depth completion, and Blender takes less than three seconds per frame of animation to render the water surface, meaning a complete 30-second video can be generated in under half an hour. This system is flexible and not tied to a particular data collection or depth completion methodology, making it an affordable and achievable approach to flood risk communication. This flexibility also means that in the future, it will be straightforward to incorporate data from more accurate photo and LIDAR sensors into the pipeline, thus further improving our results. This framework could allow for more effective risk communication that could mitigate damaging threats from future extreme flood events.