Does (element) size really matter? The effect of grid scale on flood model quality
It seems intuitive that a more granular flood map must be an improvement on a coarser resolution one. After all, fine resolution flood maps contain more detailed and precise estimates of locations that are likely to be impacted by flooding.
The truth is that whilst these maps are more detailed, this does not necessarily mean that they contain any additional skill in determining areas at risk of flooding. In fact, such precise maps can be misleading to end-users as the precision may not faithfully reflect the accuracy or uncertainty in the data underpinning them.
The most accurate widely available global terrain dataset is produced at 90 m resolution and is derived from the Shuttle Radar Topography Mission (SRTM). Finer resolution terrain data such as LiDAR data do exist in some areas and provide a more accurate alternative to SRTM-sourced data. However, the benefit of using such data is most keenly felt through its higher vertical accuracy. There is often little benefit to running flood models at native LiDAR resolution (1–5 m grid spacing) compared to when the data are aggregated to, say, 30 m resolution—particularly when you consider the quality of other components within the modelling framework: there is a limit as to how much these data can improve the model in isolation.
If a poor model – for example, one with low physical complexity in the underlying hydraulics or with inaccurately defined flow inputs – is presented at fine resolutions, it could lead to a misleading, spurious indication of precision. We have evidenced this phenomenon in the peer-reviewed scientific literature, showing that finer resolution flood models do not always lead to more accurate results—and that the influence of spatial resolution on model output pales in comparison to model parameters and boundary conditions.
The upshot of this is that a flood model is only as good as its weakest component. If your input river flows have ±40% error, what is the value of a metric-scale terrain grid? If you have no information on local drainage systems, does it matter that your elevation data is street-resolving? The components of a modelling framework need to be commensurate with one another. Modellers should focus on improving aspects of their framework to ensure that it is as close a representation of physical reality as possible, rather than focusing on producing (overly) precise models at ever finer resolutions.
A nice analogy of this is shown in the dartboard images above. In these, accuracy relates to the skill of a flood model and precision is the spatial resolution. The orange dartboard (top-left) demonstrates the best case, a highly accurate model that is also very precise. The green dartboard (bottom-left) represents a good model but one at a coarse resolution. Arguably the worst case is a model akin to the blue dartboard (top-right), whose high-precision would provide an end-user – who is likely to be unaware of model accuracy – with a false perception of the intrinsic skill of the model.
While the dark pink dartboard (bottom-right) is obviously undesirable, at least an end-user would inherently be wary of its robustness given the commensurately low-precision. Exemplifying the discussion above, the blue dartboard sells you a false prospectus on model accuracy. While the orange dartboard remains elusive at the global scale, we favour the green dartboard over any other.
You can see the accuracy–precision relationship for real-world modelling examples in images below. Here, we have modelled inundation on a tiny section (1.3 x 0.6 km) of floodplain on the outskirts of Denver, CO, USA. Although the 3 m resolution flood maps are kinder on the eye—that is all they are. They are no more skilful in the identification of flooded areas than the coarser 30 m maps, which have a near-identical flood extent. If we vary the flow inputs within their likely error margin, we can see that this has a much more drastic impact on the extent and depth of inundation. The spatial resolution had no effect on your view of risk, yet may have misled you into thinking its precision translated into accuracy. Meanwhile, the true accuracy of the flood model is illustrated by the impact of flow estimation uncertainty.
Furthermore, 10 m and 3 m resolution simulations took 13x and 412x longer to run than the 30 m data respectively—for little conceivable benefit. This enormous extra (and, largely, wasted) compute load often necessitates modellers to find other ways to speed up their simulations, for instance through simplifying the hydraulics. This can have a very real impact on the accuracy of the simulation.
Naturally, this is just one idealised example. We demonstrate the accuracy of the model through flow estimation uncertainty alone, yet there are multiple other model facets which may control accuracy too: the location and conveyance of river channels, the parameterisation of friction, or the inclusion of levees and other drainage and flow-control structures. These components hold greater sway on model performance, meaning modellers should avoid the simplistic rush to ever increasing grid granularity without necessary consideration of larger drivers of model uncertainty.
The need to be transparent about flood modelling capabilities is why we at Fathom pride ourselves in publishing our methodologies in open-access, peer-reviewed, world-leading journals. In doing so, users of our models can have the confidence that our products have a solid scientific grounding and have been derived using academically verified data and methodologies. Furthermore, as a research-led organisation, the scientific advances we herald are continually fed into our models to ensure our clients have access to products from the forefront of scientific endeavour.
Our latest global model sees a step-change improvement in accuracy thanks to the utilisation of MERIT DEM, the most accurate global terrain dataset available. MERIT DEM is a vast improvement on the previous SRTM dataset thanks to innovative algorithms that have been applied to filter out numerous sources of error. Our global model also takes advantage of the MERIT Hydro dataset published last year: the most accurate global hydrography dataset produced to date.