A 300-metre snap and a 17 percent error: what automated watershed delineation gets wrong

Click a point on a stream, wait a few seconds, and a drainage area comes back. The terrain part of that has gotten very good. I checked it against three USGS gauges in North Carolina and it landed within a few percent every time. The largest error I found came from a single setting most people never look at: how far the tool is allowed to move your outlet.

The setup

The elevation data is the USGS 3DEP 1/3-arc-second DEM, pulled from the National Map's image service in an equal-area projection (EPSG:5070). From there it is the standard chain: breach the depressions by least cost and fill whatever is left, compute D8 flow directions, accumulate flow, snap the pour point to the stream, trace the watershed, and turn it into a polygon. The area is a shoelace sum in the same equal-area projection, so there is no projection stretch in the number.

All of it runs in a web browser. The terrain tools are the Whitebox Next Gen tools compiled to WebAssembly (geolibre-wasm, MIT licensed). Nothing goes to a server except the DEM request. Each gauge took under 10 seconds end to end, and most of that was downloading the DEM.

The check is simple: pick gauges with published drainage areas, click the gauge location, and compare.

Three gauges

Cell size goes up with the basin so the window stays a manageable size. The window has to hold the whole watershed with room to spare. If the basin runs off the edge of the DEM, the area is simply truncated, and nothing warns you.

Two urban basins in Charlotte and a steep mountain basin in the Smokies, all within a few percent. That is the encouraging half.

The snap story

You never click exactly on the cell the DEM thinks the stream runs through. A gauge's published latitude and longitude, a culvert you picked off an aerial, a point you clicked at a comfortable zoom: they are all a few cells off the modeled channel. So every delineation tool snaps the point first. The rule is almost always the same. Search a radius around the click, and move the outlet to the cell with the largest flow accumulation.

To see how much that setting matters, I ran Cataloochee Creek again at several snap distances. This series was a separate run on a 20 m grid in UTM zone 17N, from the same 3DEP source:

Nothing from 0 to 150 m. Then 8.5 square miles appear at 300 m. The gauge sits just upstream of where Little Cataloochee Creek comes in. Inside a 300 m radius, the cell with the most flow accumulation is no longer on Cataloochee Creek at all. It is on the combined stream below the confluence, which carries both creeks. The tool did exactly what it was told. It found the biggest number in the circle, and the biggest number belonged to a different watershed.

This is not a Cataloochee quirk. Gauges, culverts and road crossings often sit a short distance above a confluence. Every one of those is a place where a generous snap radius will hand you the wrong watershed. And the result does not look wrong. The polygon is clean, the boundary follows the ridges, and the area is only 17% bigger, which is well within what you might believe for an unfamiliar basin.

Why the urban basins matched, and why you should not count on it

Little Sugar Creek drains a good piece of central Charlotte, and it came out 2.4% high. I would not read that as proof that terrain delineation works in cities. The DEM sees streets, curbs and swales. It does not see a 48-inch pipe that carries a neighborhood under a ridge into the next basin. At 11.8 mi², a few of those crossings wash out.

For site-scale work, under about 50 acres, a 3DEP delineation is a starting sketch. The real boundary comes from survey or lidar plus the storm drain plans, and an engineer who has looked at both.

A tooling note for the curious

While building this I hit a quieter version of the same problem. A widely used open-source D8 accumulation tool, when fed a 16-bit flow-direction grid, writes its output in that same 16-bit type, so accumulation caps at 32,767 cells:

Past about 3.3 km² at 10 m cells, every cell on a big stream reads the same number. On a 30 m test grid, 238 cells sat at exactly 32,767 while the true maximum was 93,573. A snap that looks for the largest value in its radius now finds a tie instead of one main-channel cell, and where the outlet lands is down to the tool's tie-breaking. The workaround is to accumulate from the DEM itself rather than from the pointer grid. I have reported it upstream. The general lesson is older than this tool: when a grid of counts has a suspiciously round maximum, check the data type before you check the hydrology.

Practical rules

• Keep the snap radius to a few cells. At Cataloochee anything up to 150 m (7 or 8 cells at 20 m) was safe, and 300 m was not. If you need a bigger radius to land on the stream, the click is in the wrong place.
• Look at the outlet against the stream network, not just at the polygon. If the snapped point is below a confluence you meant to stay above, the area is wrong no matter how clean the boundary looks.
• Check the area against something independent when you can: StreamStats, or a USGS gauge on the same stream. A gauge's published area is the best free check there is.
• Make sure the window holds the whole basin. A watershed that touches the edge of the DEM is truncated.
• In town, find the pipes. The DEM delineation is the surface answer. The storm drain system decides the real one.

Where this lives now

HydroComplete's watershed inspector has a Delineate from Map button that runs this pipeline in your browser. Click the outlet, look at the polygon and the snapped point, and the area goes into the node. Because it is a screening number, the report marks the area [3DEP screening, unverified] until you replace it with one you have checked. That tag is deliberate. A drainage area that came from a click should say so on the page the reviewer reads.

— Michael Flynn, PE
Next issue: back to the pond outlet for its third personality — when the riser barrel itself becomes the control, and how to tell weir, orifice, and pipe control apart from the stage-discharge curve alone.