Development of a fusion methodology for flood mapping using SAR and optical satellite data

In this thesis, a feature-level data fusion methodology was developed using the random forest (RF) approach, which is one of the frequently utilized machine learning methods for natural hazard assessments. For this purpose, Sentinel-1 SAR and Sentinel-2 optical data provided regularly and free of charge by the European Space Agency were used. For the development and performance analysis of the method, two study areas/flood cases with different topographical characteristics and data availability conditions were analyzed. The first case is a dam failure in Uzbekistan-Sardoba, for which pre- and post-event Sentinel-1 and Sentinel-2 and external reference data (PlanetScope) are available. Since the flood disaster in this region was not caused by precipitation, it is independent of the cloud effect that obstructs the Sentinel-2 data after the event. In addition, the well-known synthetic aperture radar (SAR) distortions that disturb the Sentinel-1 data could be omitted thanks to the flat topography of the area. For this reason, the method proposed in the thesis was primarily developed in Sardoba and the results were evaluated by applying it to another area, the Türkiye-Ordu May 2018 floods. The Ordu flood case was caused by precipitation, and unlike the first study area, the region has rugged topography. When the results obtained were validated with external data, it was observed that especially flood and flooded vegetation classes could be determined with high accuracy. In the thesis, flood mapping strategies to be undertaken based on data availability scenarios are discussed in detail.