Online Anomaly Detection on Rain Gauge Networks for Robust Alerting Services to Citizens at Risk from Flooding

The modern cities are addressing their innovation efforts for facing not just the common stresses cities accumulate daily, but also the sudden shocks can occur such as urban floods. Networked gauge stations are instrumental to robust floods alerts though they suffer from error and fault. For capturing the anomalous behavior of networked rain gauges, the use of an online anomaly detection methodology, based on the Support Vector Regression (SVR) technique, has here been investigated and developed. The specific anomaly case of incorrectly zero sensor readings has been efficiently addressed by a centralized architecture and a prior-knowledge free approach based on SVRs that simulate the normality profile of the networked rain gauges, on the basis of the spatial-temporal correlation existing among the observed rainfall data. Real data from the pilot rain gauge network deployed in Calabria (South Italy) have been used for simulating the anomalous sensor readings. As a result, we conclude that SVR-based anomaly detection on networked rain gauges is appropriate, detecting the eventual rain gauge fault effectively during the rainfall event and by passing through increased alert states (green, yellow, orange, red).