This is somewhat of a theoretical question stemming from some discussions with colleagues on the topic of implications with delineating basins with projected (e.g., Albers Equal Area) vs. unprojected (NAD 83) data derived from a 10m DEM that's in NAD 83.
Some have stated that it's not an issue as the values calculated from unprojected data simply get adjusted if you do decide to project.
I'm not sure this is the case though, as there are inherent differences between data in a geographic coordinate system and projected data. I tried one example going through the routine starting with unprojected DEM data, then tested the same site with projected DEM data. Steps taken for both were done (all work done in ArcGIS 9.3.1) using 10m DEM data.
One run was done using a DEM in NAD 83, and the second run was done by projecting the same DEM into USA_Contiguous_Albers_Equal_Area_Conic_USGS_version.
- derive flow direction using the geoprocessing FlowDirection_sa tool
- derive flow accumulation using the FlowDirection_sa tool
- snap the pour point using a 50 meter distance
- delineate the watershed using the Watershed_sa tool
In comparing the two I could notice a visual difference between the display of the Flow Direction grids.
NOTE: After more subsequent research I believe the striping effect is due to not using a CUBIC resampling but by mistakenly going with the default of NEAREST in the ArcGIS Project Raster tool. I don't believe this provides any sort of resolution to this debate though...
Flow directions using unprojected DEM
Flow directions using projected DEM
I understand that visual comparison is not 100% scientific but can be a good starting point.
Accordingly, there was a difference between the pour point with how it snapped for each run. And, there was a definite difference in the derived watersheds given how the snap pour point tool decided to snap based on the respective projected/unprojected datasets. The watershed shown in green is the watershed derived using the projected DEM and subsequent projected-derived elevation derivative data. The watershed shown in the purple outline is the watershed derived using the unprojected DEM data.
I've come across these two other GIS forum threads (links below) that discuss this issue in the old ESRI forums, but I'm still not clear as to how the Flow Direction tool works relative to projected vs. unprojected data (I understand the concept of hydrologic flow and flow direction though). If each cell still has the same elevation value in a projected DEM vs. an unprojected DEM (is this correct?), why is there a difference in a flow direction raster derived from projected data versus one derived from DEM data in NAD83?
Also, would any differences theoretically be less of an issue if doing delineations in a higher Latitude such as, Shenandoah National Park in Virginia versus doing delineations in the state of Texas?
I spoke with one cartography expert that thought that the east-west distortion you get as you move away from the equator could likely be an issue (like how in some maps Canada is extremely bloated and distorted), in that if you're more than 10 degrees of latitude away from the equator they thought projected data is the way to go if you're concerned with accuracy.
One major unknown is the level of uncertainty with basins delineated using unprojected data that we're trying to get a handle on. There is a difference, but what is the magnitude?
Thanks to anyone that can provide a straight-forward answer to this discussion, or just some helpful insight into this.
The main issue we're interested/concerned with is if there will be accuracy issues with the delineated watersheds as a result of starting the process using an unprojected DEM.
So, if I'm understanding the reply, the delineated basins should be fine in terms of representing the drainage area for a pour point? It seems though if the flow directions are wrong that will result in some error in the final delineated watershed.
This is a very interesting and really an important topic - I have yet to see a report or documentation stating it's OK to use UN-projected data for delineating watersheds. I have set through ESRI User Conference technical talks led by the lead developer engineer of the Spatial Analyst extension (which houses the Hydrology tools) where they said you should use an equal area projection (such as Albers equal area) as well.
As well, there doesn't seem to be any authoritative "bible" standard for how to go about this - just seems that it's an almost acknowledged de facto approach to project the data before calculating your elevation derivatives.
Nowhere have I been able to find a concise and straightforward answer as to how this impacts flow direction calculation and subsequently the delineation of a watershed.
And, if you end up working with watersheds delineated using unprojected DEM data and then you project those watersheds, isn't the inaccuracy still there (e.g., in terms of determining a watershed area or any other characteristics such as land cover proportions etc)?
Furthermore, I'm assuming that projecting a flow direction raster that was derived from an unprojected DEM does not correct the errors either since the source data were unprojected....
thanks - appreciate any additional insight you can provide
EDIT - 20110331
thanks for this extensive discussion. We've been researching this issue more and actually have come across some references that suggest that it's actually better to not project the DEM before getting flow dir., flow accum., and delineating.
One email response from an anonymous source (but who is a pretty reputable person), when posed the question of 1.) project DEM 2.) produce derivatives OR 1.) produce derivatives 2.) project DEM said:
In a nutshell, it depends on the derivative. For continuous derivatives that will be visualized, you should derive and then project—this reduces the risk of tile boundary artifacts being enhanced or introduced (by the projection algorithm) and then passed along to the derivative if you were to project the DEM first. The exception to this is when you are also using distance or area as the basis for your derivative calculation. This is of course relative to how large the distances/areas are and how far you can acceptably get away from the equator. So imagine that for derivatives like slope or hillshade, which depend on the cellsize, there are consequences. These derivatives will be most accurate at the equator, and the accuracy will degrade significantly past 60 degrees north or south. In both cases, I am assuming the DEM covers a very large area (wider than 1.5 UTM Zones) and a traditional tile-based approach where the tiles are either arbitrary or conform to existing standards like USGS Quad sheet boundaries. So saying the implication is that much of this thinking predates mosaic datasets, which I am less able to comment on. The main concern for me would be wanting to know how well matched the DEM tiles are. If they are well matched (like NED) then I expect things to work well, with derivatives being derived from tiles (as functions applied to the mosaic dataset) and then these are displayed on the fly. If they are not well matched, then garbage in, garbage out. Back to your original question, I think if it is just watershed boundaries, it would be possible to derive these without projecting because it’s not how much curvature or slope that matters, just where it is and that it exists.
They went on to say:
The reason I would stick to the un-projected methodology is that we are using rasters which are in and of themselves a derivative of DEM (which we typically don’t have, but think LiDAR point cloud). For rasters that cover very large areas, like continents at relatively fine levels of resolution, projecting to something like Albers will result in loss or introduction of information, when the raster uses regular sized cells (like Esri’s rasters do). That means tools like Flow Accumulation will produce results based on partial or interpolated information. Basically all projection algorithms applied to rasters will cause problems as soon as there is a expansion or shrinkage of more than the distance of a pixel width (projections like Albers can introduce error by introducing new pixels between two old ones). Deriving from these means the potential for cumulative error is high.
This seems to suggest the opposite - that projecting introduces more noise, unless you get above 60 degrees latitude.
We've also come across some published sources that have suggested similar that unprojected is an acceptable approach for smaller watersheds (last 2 paragraphs of section 1.6) from Distributed Hydrologic Modeling for GIS (Vieux, 2004): http://www.springerlink.com/content/x877238532533g20/fulltext.pdf
So, in the end, does it just boil down to a matter of 1.) where you're doing work on the earth's surface 2.) the scale you're working at, and 3.) whether the noise introduced by a projection that will better preserve attributes that affect the flow direction algorithm is less than the distortion introduced by unprojected data (the benefit increasing as you move towards the poles) to determine whether you should project to something like conformal, or if it doesn't matter?
When you start digging into this topic it seems like the larger consensus is to project, but there are some that seem to say that's not a hard and fast rule.