Closing the Gap: Towards Consistent Flood Extent Retrieval with Multi-Sensor Data Fusion

Despite the significant increase in Earth Observation (EO) satellites, the frequency of cloud-free imagery at sufficiently high spatial resolutions for timely inundation mapping remains a significant challenge. Obtaining more frequent flood extent estimations would contribute to our understanding of flood dynamics and increase the likelihood of capturing the flood peak, which often evades EO acquisitions. Integrating complementary data from multiple sensors is a potential solution to overcome limitations posed by temporal resolution, spatial resolution, cloud cover, adverse weather, or light conditions. Surface water fractions, indicating the proportion of a pixel covered by water, can be derived from a variety of sensors that passively sense across different spectral ranges daily. However, the fractional coverages are derived at various spatial resolutions, necessitating a methodology to harmonize and combine the information to obtain a comprehensive flood map at a meaningful resolution. The present study proposes a methodology to seamlessly combine data from Low-Earth Orbiting (LEO) multispectral, Geostationary-orbiting (GEO) multispectral, and Passive Microwave (PMW) sensors. The proposed approach is tested on the February 2022 flood event in Brisbane, Australia, and fuse data from Visible Infrared Imaging Radiometer Suite (VIIRS), the Himawari 8/9 Advanced Himawari Imager (AHI), and the Special Sensor Microwave Imager/Sounder (SSMIS). These sensors offer complementary strengths in flood detection, including sub-daily imagery from VIIRS and AHI, and fractional water estimates beneath cloud cover from SSMIS.

Surface water fractions, representing the fraction of a pixel covered by water, are derived from VIIRS, AHI, and SSMIS at spatial resolutions of 375 m, 1 km, and 25 km, respectively. These surface water fractions are subsequently homogenized via downscaling and fused to obtain an aggregated flood map. A Digital Terrain Model and its derivatives, including the Slope, Topographic Water Index, Height Above Nearest Drainage, and Flow Accumulation, and water frequency information are utilized to downscale and distribute the surface water fractions in physically plausible ways. This disaggregation process produces comparable flood maps from all sensors. These maps are thereafter combined to yield a single detailed flood map. This multi-sensor framework ensures the consistent generation of flood maps at a meaningful spatial and temporal resolution, compensating for the unavailability of moderate- to high-resolution imagery due to satellite revisit timing and cloud obstruction. The proposed approach enables more frequent generation of detailed flood maps, providing valuable insights into inundation dynamics to scientists and decision makers.