Earth Observation Data Fusion for rapid flood extent mapping in data-limited areas: A Case Study of Tana River Delta, Kenya

Floods account for more than 40% of recorded natural disasters in the last two decades (EM-DAT). Given the current climate dynamics, flooding is likely to increase in intensity with shorter return periods. Flood risk is highest in the Global South countries due to lack of adequate prevention, prediction and monitoring systems.

However, the evolution of earth observation technology over the last decade carries a great potential for accurate, near real-time flood extent mapping and prediction. While this favors synthetic aperture radar (SAR) based approaches due to its ‘all weather’, day or night and canopy penetrating capabilities, the flooded scene presents a complex environment characterized by mixed signal to surface interaction mechanisms that complicate SAR imagery analysis. This necessitates fusion with other remote sensing flood detection techniques.

This study aims to develop an earth observation data fusion model for near-real time flood detection, early warning and risk as well as damage assessment in data-limited areas that fall within the exclusion mask of the Copernicus Global Flood Awareness System (GloFAS). The area of study is Tana Delta Ramsar Site in Kenya, a productive ecosystem comprising of a unique mix of fresh water, floodplain, estuarine areas and beaches supporting several ecosystem services. This analysis focuses on the April 2024 flooding event.

The initial step involved comparative analysis of three Sentinel-1 (S1) flood detection techniques namely image segmentation (Otsu threshold), multi-temporal change detection (CD) and a hybrid (Otsu + CD) technique, post-processed for removal of permanent water as well as slope and spatial-context conditions. With the limitation of missing validation data, Sentinel-2 (S2) image classification was used albeit with a 3-day acquisition date misalignment between the post-flood S1 and S2 images thus assuming no significant land cover changes took place.

Preliminary results show higher flood areas detected by the VH vis-à-vis the VV channel. In particular, the Otsu detected 240.85 km² for VH and 196.98 km² for the VV while the CD returned 40.64 km² and 27.56 km² for the VH and VV respectively with a change threshold of 1.5. Lastly, the hybrid approach detected 143.66 km² for the VH and 47.33 km² for the VV against S2’s 194.29 km². This difference could be due to the depolarization of the VV backscatter in the vegetated areas.

The next steps will involve testing of operational workflows such as the Copernicus Global Flood Monitoring (GFM), UN-SPIDER recommended flood detection approach and the AUTOWADE 1.0 (AUTOmatic Water Areas DEtector) in a data limited context (Tana Delta, April 2024 flood event). Further, the Otsu threshold required initialization of a bimodal histogram for best performance while the CD results rely on the chosen change threshold hence are not suitable for automated inundation mapping. Consequently, the study will explore the use of AlphaEarth and Clay geo-foundation models for rapid flood mapping in complex land use/ land cover and data-limited areas.

Key words: SAR, Data fusion, Flood extent mapping, geo-foundational models