Problem: Flooding Without Warning How to predict river flooding 24 hours in advance given: • Large geographic areas of 10000 km 2 • Limited sensing infrastructure • Difficult environmental conditions • No existing infrastructure • Numerous contributing variables and no historical data • Computationally extensive and centralized prediction models Solution: Regression Prediction Model Solution: Sensor Network for Autonomous Early Warning System Architecture • 900 MHz mini-networks of sensing nodes measuring precipitation, temperature, and river pressure with support for additional sensors if needed • 144 MHz radio backbone running the distributed computation model to determine flood risk • Office nodes to provide flood alerts, to assist in maintaining system, and to obtain Internet satellite and weather data • Community nodes to alert of incipient flooding Field Experiments: Honduras and Dover, Massachusetts Rio Aguán Basin Tocoa, Honduras Tocoa Field experiment of system in Honduras: • Installed: - 4 sensing nodes: 2 precipitation and 2 temperature - 2 computation nodes with river level sensors - 1 office node • Computed distributed calibration and prediction model Field experiments in Charles River at Dover, Massachusetts: • Fall 2007 - Installed 3 sensing nodes; one of each type - Gathered 4+ weeks data • Fall 2008 - Installed 5 sensing nodes: 1 pressure, 2 temperature, and 2 precipitation - Ran distributed prediction with predetermined coefficients - Operational for approximately 2 months Installed infrastructure in Honduras: • Radio antenna towers: - 5 meter for river locations - 10 meter for offices • Solar power backup system for offices with automatic switching from grid • Pressure sensor installation (bridge design only) Model‐ Based Monitoring For Early Warning Flood Detection Elizabeth Basha and Daniela Rus Distributed Robotics Lab • Calibrate model using novel 3-step distributed pseudoinverse algorithm where A is mxn and m≥n: 1. QR Decomposition: [Q, R] = qr(A) 2. Singular Value Decomposition: [U, R, V] = svd(R) 3. Pseudoinverse Combination Step: x opt =VR -1 (QU) T b • Verified in Matlab using randomly generated matrices • Implemented on sensor network and tested using internal temperature sensors; system ran error-free for 8 hours with calibrating the model every 10 minutes Other Applications • Predict congestion in networks of multi-function devices that can provide print, scan, fax, and email service along with other functionality in office environments (with Xerox) • Predict water usage across multiple agriculture fields in order to provide efficient and dynamic control of water irrigation systems (with CSIRO Brisbane) • Need simple prediction computation that can run distributed on sensor network • Developed multiple linear regression model operating on locally sensed river level, precipitation, and air temperature data • Verified using 7 years of data from the Blue River in Oklahoma: - Sensors consist of 6 precipitation, 1 air temperature and 1 river flow - Model uses 1 year of data for training Precipitation Sensing Node Computation Node Pressure Sensor