We've explored several new strategies. First, we've developed a new AI method that uses satellite data from Sentinel-2 to create more accurate flood maps from the VIIRS sensor. This new method is about twice as precise than previous techniques used by NASA for major flood events.

Second, we’ve shown that using a new, publicly available dataset called FloodPlanet, which includes detailed flood information from commercial satellites, helps improve the accuracy of flood maps generated from other satellites like Sentinel-1 and Sentinel-2 by 15%.

Finally, we've looked at whether very advanced AI models, called 'geospatial foundation models,' are better at mapping floods than more common AI methods. We found these advanced models (testing Clay, Privthi, and DOFA in this experiment) offer slightly better overall accuracy when a lot of training data is available. More importantly, when there's only a small amount of training data, these advanced models make a big difference, producing much more