Time-series analysis of Sentinel-1/2 data for flood detection using a discrete global grid system and seasonal decomposition

Automated flood detection using earth observation data is a crucial task for efficient flood disaster management. Current solutions to identify flooded areas usually rely on calculating the difference between new observations and static, pre-calculated water extents derived by either single acquisitions or timely aggregated products. Such pre-calculated datasets, however, lack representation of real-world seasonality and short-term changes in trend. In this paper we present a complete workflow to automatically detect hydrological extreme events and their spatial extent, which automatically adapts to local seasonality and trend. For that we rely on a novel combination of well-established algorithms and tools to detect anomalies in time-series of water extent across large study areas. The data is binned into a discrete global grid system H3, which greatly simplifies aggregation across spatial and temporal resolutions. For each grid cell of an H3 resolution we perform a time-series decomposition using Seasonal and Trend decomposition using Loess (STL) of the cell’s proportion which is covered with surface water. All cells receive an anomaly score, calculated with extended isolation forest (EIF) on the residuals for each step in time. A burst of anomalies represents a hydrological extreme event like a flood or low water level. The presented methodology is applied on Sentinel-1/2 data for two study areas, one near Sukkur, Pakistan and the other one in Mozambique. The detected anomalies correlate with reported floods and seasonal variations of the study areas. The performance of the process and the possibility to use different H3 resolutions make the proposed methodology suitable for large scale monitoring.