Flood Detection using Semantic Segmentation and Multimodal Data Fusion
Bipendra Basnyat, Nirmalya Roy, Aryya Gangopadhyay
Abstract
Real-time flood detection and notifying the citizens about its risk is of utmost importance. This work discusses the real-time deployment of one such notification system called Flood-Bot. FloodBot is a vision-powered flood detection and notification prototype deployed in a flash flood-prone Ellicott City, Maryland. We discuss the real-time deployment of FloodBot and our approach in detecting the flood event using semantic segmentation and multimodal data fusion. We implement the state-of-the-art semantic segmentation model U-Net and its modified version to track landmass with an accuracy of above 80%. We augment the parsed scene data with actual flood level sensor readings and ambient weather data for better scene representation. We validate the deep learning model's outcome using the flood sensor before posting risk message into social media. We then articulate the learning and challenges around our deployment from June - November 2020.