dem - When computing flow direction, the D8 algorithm is used - are there major differences if others are used?

I am a student working with Hydrological models. Within these models we can route water across a DEM (within a delineated watershed or just within an un-delineated DEM) using many different flow routing algorithms: D8, MFD (sometimes FD8), Rho8, DEMON, D-Inf, and so on, with different variations within even the ones I listed possible (original MFD vs. MFD with a power term).

But, when working within a delineated watershed, a flow routing algorithm (primarily the D8) has already been applied to get the flow directions grid - typically the first step in a delineation.

Has there been any research looking at the difference between delineations performed on flow directions grids created using the other algorithms? I know since these other algorithms allow multiple flow directions, that it might be difficult to visualize this data (unless there is some compounding into one value) - and that that alone might make it impossible.

It doesn't seem like the delineations would be any different - if point y is in the watershed of pour point x, water will flow from point y to pour point x, whichever way it may get there. The flow may be split into an infinite number of paths (infinite dispersion), but they will still all end up at point x, right? So even though D8 might not be perfect for predicting actual flow paths, it should produce the same delineation as any other, right?

Answer

It may seem like laziness on the part of Watershed tool developers to stick with the simplest and oldest flow algorithm, D8, but there is a very sound reason for doing so. The difference between the D8/Rho8 flow algorithm and the more advanced algorithms that you mention (e.g. D-infinity) is mainly in their inability to represent the dispersion of overland flow on hillslopes. D8, Rho8 and D8-LTD (Orlandini's method) are each non-dispersive algorithms, i.e. each grid cell in the DEM will contribute flow to only one downslope neighbour. Viewed another way, it's like saying that there is only one unique flowpath issuing from each grid cell. Non-dispersive flow algorithms are particularly suited to watershed delineation for this reason. With a dispersive flow algorithm, grid cells near the boundary of a watershed can actually belong to both watersheds, i.e. two of the multiple flowpaths that issue from a grid cell situated near a watershed boundary may actually terminate at different outlets entirely. In reality of course watershed boundaries are somewhat fuzzy and ambiguously defined, but we generally want delineated watersheds that are mutually exclusive. Since the fuzziness caused by the flow dispersion inherent in the flow algorithm isn't necessarily related to the physical processes that actually govern the ambiguity of flow near watershed boundaries, we have no basis for interpreting the dual membership of bordering grid cells. That is, we can't say, 'this grid cell belongs to this watershed 90% of the time and its neighbouring watershed 10% of the time' solely based on the artificial nature of the dispersion imposed by the algorithm design. Dispersion is particularly problematic therefore. I hope that clarifies the reason why most Watersheds tools are based on D8 (though really Rho8 and D8-LTD would be equally suited).