Supporting flood disaster response using multi-sensor Earth Observation data

Abstract

Remote sensing data has become an essential part of today's crisis management activities. In recent years, the German Aerospace Center (DLR) has developed various components to support flood disaster response using multi-sensor Earth Observation (EO) data. On a global level, a multi-sensor system for automatic and large-scale surface water extraction was implemented. The system consists of several cloud-based modular processing chains based on convolutional neural networks (CNN) to extract the surface water extent from systematically acquired high-resolution radar (Sentinel-1) and multi-spectral (Sentinel-2 and Landsat) satellite data. A globally applicable high-resolution seasonal reference water product at 10-20 m spatial resolution based on fused Sentinel-1/2 time-series data over a reference period of two years is computed and used to distinguish permanent water from temporary flooded areas. The system also allows to provide information about the duration of flood coverage on a pixel level by combining single temporal flood masks over time. Further, a mechanism has been installed to identify whether the water extent outlined in a satellite scene is abnormally large or small in comparison to a reference period. The anomaly detection criterium is based on the interquartile range (IQR). In case of observed anomalies, end users are alerted using E-Mail notifications. To enhance situational awareness, early-stage estimations of impacted regions derived from heterogenous geospatial indicators can aid to prioritize the crisis management activities and support data collection initiatives of very high-resolution (VHR) EO imagery (satellite, aerial, UAV). In this context, a log-linear pooling method coupled with an unsupervised hyperparameter optimization routine is developed to fuse information on flood hazard extracted from high-resolution satellite imagery with disaster-related data from geo-social media and freely available supplementary geospatial data on exposed assets (e.g. building distribution, population density, hazard zones). The identification of disaster hot spots is carried out on the basis of the H3 global grid system. Very high-resolution EO data tasked in the frame of an activation of a crisis mechanism on-demand and supported by rapidly generated disaster hot-spot maps, are analyzed using deep learning-based approaches in the frame of the multi-sensor EO system. The spatial resolution of these sensors enables the identification of relevant local crisis information, e.g. small-scale flood extent in heterogenous landscapes as well as damaged buildings and infrastructure to provide a detailed and reliable picture of flood-affected areas in order to optimize disaster management activities.