Pluvial flooding assessment using remote sensors and object detection models

In this work, two different remote sensing technologies were employed to support the assessment of pluvial flooding scenarios: the Smart Rainfall System (SRS) to estimate the rain rate and aerial photos for object detection purposes. The SRS is a recently developed monitoring technique able to estimate the rainfall intensity by processing the attenuation of microwave signals from satellite links measured by low-cost sensors. To accurately identify exposed objects to flood hazard, advanced object detection algorithms based on deep learning techniques are employed. The proposed methodology was applied to a case study located within the metropolitan area of Genoa (Italy), characterized by a flat area of about 1 km² and recently affected by a pluvial flooding event characterized by rainfall intensities having a return period lower than three years.

The study area is equipped with a traditional tipping-bucket rain gauge station and one SRS. Two further SRSs and two rain gauges are available close to the investigated area. This configuration allows to mimic different rainfall monitoring levels from the ungauged basin to a higher spatial resolution. Pluvial flooding scenarios were modelled using the HEC-RAS 2D software and results show that significant differences in the expected flood volumes and maximum water depth and velocity are obtained using various sources of rainfall data. The obtained differences reveal that the role of opportunistic sensors located within or in the proximity of the study area largely outperforms the contribution of nearby rain gauge data when these are located even only 5 km far from the study area. This is ascribable to the point nature of measurements taken by rain gauge against the more spatially extended rainfall estimation provided by the SRSs. 

The object exposed to flood hazard were detected using the You Only Look Once (YOLO) models applied to aerial images at a spatial resolution of 5 cm. The performances of various YOLO models were investigated. YOLO is pretrained for the detection of vehicles while samples of the aerial images selected outside of the study area were used to train the model for the detection of trash-bins. Results for vehicles and trash-bins are characterized by an accuracy of 95% and 69%, respectively. The assessment of the accuracy of the model based on the spatial resolution and the presence of shadows is still ongoing. Results will allow to assess the vulnerability of the urban context and will be combined with the flood hazard maps to obtain flood risk scenarios. 

This study was carried out within the RETURN Extended Partnership and received funding from the European Union Next-GenerationEU (National Recovery and Resilience Plan – NRRP, Mission 4, Component 2, Investment 1.3 – D.D. 1243 2/8/2022, PE0000005).