FLORAS: urban flash-flood prediction using a multivariate model

Hydrological models allow water levels to be predicted at critical spots when the problem of flooding is being addressed. However, these models fall short in their attempts to provide timely warnings to communities at risk as they often involve complex setup requirements and incur high computation costs. Other approaches have been adopted that make use of water level monitoring sensors for detecting floods. Although accurate in their performance, these approaches often require a high level of maintenance because their predictions rely on critical readings from sensors that have to be immersed in rivers. We recommend a machine learning-based methodology for flood detection to address this issue, called FLORAS. It makes it possible to build models that make predictions solely on the basis of meteorological data from weather stations-water height measurements are only needed to employ ground truth for the purposes of training and validation. We evaluated the methodology with current data readings from São Carlos (SP - Brazil) in experimental analyses. Water height measurements from sensors placed at sites along the river were correlated with open weather data from a reputable, local source (Climatempo – Brazilian weather). The results show that the model achieved a higher degree of of accuracy and incurred lower computational costs than SwMM, a hydrological model. These results show that the recommended methodology is suitable for systems that run with resource-scarce devices, such as the IoT systems that are usually deployed in flood detection frameworks.