Global Flood Models (GFMs) and Earth Observation (EO) play a crucial role in characterising flooding, especially in data-sparse, under-resourced regions of the world. However, validation studies are often limited to a handful of historic events and do not directly assess the ability of these products to simulate flood hazard - the probability that flooding will occur in a given location. As a result, it is difficult for stakeholders to decipher the ability of either models or observations to identify flood hazard and make decisions to mitigate for flooding. Here, we leverage flood observations from 20 years of MODIS data to compare the recorded flooding with what would be expected given the hazard simulated by a GFM. We devise an approach, Flood Expectation Per Pixel, and apply it across four large basins in Africa – Congo, Niger, Nile and Volta representing a variety of biomes. We estimate the uncertainty of EO to capture flood events due to burned areas, cloud cover and vegetation, incorporating uncertainty estimates when comparing to modelled hazard. We found that at lower return periods (<20 years), the EO data records less flooding than the GFM, suggesting GFMs overpredict frequent flooding. For return periods between 50-100 years), GFM and EO data show greater consistency given the uncertainties we consider. For large return periods (100 years) the EO observations show more flooding than expected given the GFM data, potentially due to data errors and non-fluvial flooding, however there are too few observations to draw significant conclusions at these return periods. The EO record indicates that the GFM can differentiate between flood return periods. We find EO and GFM complement each other and thus should be used in tandem to inform strategies to mitigate floods across the hazard spectrum from frequent to extreme flood events.