Rapid Building Damage Detection Through SAR Timeseries Analysis in the Google Earth Engine

Abstract

Humanitarian crises related to building and infrastructure damage, i.e. natural hazards and collateral damage during warfare, result in many hundreds of thousands of casualties on a yearly base. Earthquakes, for example, are responsible for taking around 50.000 lives on a yearly base due to falling debris and collapsing buildings, making them one of the most deadly natural disasters known to mankind. These casualties are caused by both direct effects (falling debris) as well as indirect effects (being trapped by debris without access to air/water/food). In the first hours after an earthquake hit an area, a clear overview of the situation is hard to obtain due to inaccessible roads and damaged communication systems, which in its turn hampers setting up rapid disaster response operations. In order to assist the search and rescue teams in wide-area situation awareness, remote sensing techniques can be used to acquire and create damage detection maps which could be used to prioritize their tasks and help save more lives.

Spaceborne Synthetic Aperture Radar (SAR) is one of the most promising techniques in rapid disaster response as it is an active sensor which remains unaffected by the weather, smoke, cloud cover and/or daylight. The Sentinel-1 mission is a relatively new SAR mission that acquires data over the entire Earth's surface every 3 days in any viewing geometry. More importantly, this mission has an open-data policy and hence provides the opportunity to easily perform multi-temporal data analysis instead of the bi- and tri-temporal change detection approaches proposed in previous studies. The development of a fully automatic change detection algorithm has been done in the Google Earth Engine: an online, open-source processing tool that allows users to run algorithms on geo-referenced Earth observation imagery stored on Google's infrastructure, with one of these datasets being the global Sentinel-1 GRD dataset. The time reduction that results from bringing the algorithm to the data instead of vice versa as well as the possibility to easily combine various geo-referenced datasets both benefited the rapid disaster response operations.

The processing of the data is performed completely automatically based on the date and location of the disaster as input, making it easy to use for non-remote sensing specialists. Using these two parameters, the Sentinel-1 images overlapping the area of interest are selected after which a change detection algorithm is applied to the timeseries. The resulting damage detection maps has been validated in terms of operability and accuracy for wide-area situation awareness in search and rescue operations for three case studies: the 2016 M6.2 Central Italy earthquake, the 2017 M7.1 Central Mexico earthquake and the 2018 Syria military strikes. Overall, the algorithm performs well in terms of damage detection, with a low amount of false positives and an even lower amount of false negatives, for both damage due to natural hazards and intentional or collateral damage during warfare. The algorithm experienced difficulties when identifying individual collapsed buildings in a dense urban environment as around 50\% of the collapsed buildings were detected as such. However, the accuracy of the damage detection is much better when looking at clusters of damaged buildings in urban areas or at individual buildings, which are equal to or larger than the size of a SAR resolution cell, in less dense urban environments.

The complete approach and final damage detection maps have been presented to the Urban Search And Rescue team of The Netherlands, whom are the potential end-users of these maps. They saw great potential in this approach in general to help prioritize their search and rescue operations, with the only bottleneck being the data latency. We showed that once a new image has been acquired after the disaster, the total data latency by ESA and the GEE is at most 72h. This is also the maximal time the USAR team has in order to move from their home base to any country in the world in case of a disaster. The data latency time frame is a maximum, so actual averages are lower and the final damage detection map will be completed and ready to use once needed by the USAR team. This illustrates that the approach proposed in this research can be used on an operational base to assist in rapid disaster response situations.