Predictive multi-watershed flood monitoring using deep learning on integrated physical and social sensors data

Abstract

This paper presents a deep learning model based on the integration of physical and social sensors data for predictive watershed flood monitoring. The data from flood sensors and 3-1-1 reports data are mapped and fused through a multivariate time series approach. This data format is able to increase the data availability (partially due to sparsely installed physical sensors and fewer reported flood incidents in less urbanized areas) and capture both spatial and temporal interactions between different watersheds and historical events. We use Harris County, TX as the study site and obtained seven historical flood events data for training, validating, and testing the flood prediction model. The model predicts the flood probability of each watershed in the next 24 hours. By comparing the flood prediction performance of three different datasets (i.e., flood sensor only, 3-1-1 reports only, and integrated dataset), we conclude that the physical-social data integrated approach can better predict the flood with an accuracy of 0.825, area under the receiver operating characteristics curve (AURC) of 0.902, area under the precision-recall curve (AUPRC) of 0.883, area under the F-measure curve (AUFC) of 0.762, and Max. F-measure of 0.788.