The places with the least data are the places with the most flooding
Most flood modeling effort goes where the observations already are. That is understandable, and it is backwards.
Flood modeling has a distribution problem. The coastlines with the densest gauge networks, the best bathymetry, and the longest records are also the coastlines with the most modeling attention. The places where a warning would change the most lives have the least of all three.
This is not anyone's fault. A model is easier to build, easier to defend, and easier to publish where it can be validated. Reviewers ask for skill statistics, and skill statistics need observations. The incentive points away from the basins that need the work.
What "data-scarce" actually means
It rarely means no data. It usually means:
- A gauge network that is sparse, intermittent, or was destroyed by the last event
- Bathymetry interpolated from charts surveyed decades ago
- Terrain from a global DEM with vertical error comparable to the flood depth
- A rainfall record too short to say anything confident about return periods
Each of those is a real constraint. None of them is a reason to produce nothing.
Building around what exists
Global datasets have improved enormously. Reanalysis gives consistent meteorological forcing anywhere on Earth. Global bathymetry and satellite-derived terrain cover the domains that surveys never reached. Satellite radar observes flood extent through cloud, which is exactly the condition during the event you care about.
None of that is as good as a well-instrumented site. All of it is dramatically better than nothing, and the gap between "uncertain model" and "no model" is far larger than the gap between "uncertain model" and "good model".
The part that has to come with it
Working this way puts more weight on stating what the result can and cannot support. A depth grid built from a global DEM is not a first-floor elevation survey and must never be presented as one. The honest output is a result plus an explicit boundary around it.
The alternative to an uncertain model is not a better model. It is no warning at all.
That is the trade we think is worth making, and the reason the limits of each method are published alongside the method.