I have a point dataset expressed in British National Grid (about 10,000 points distributed all over the UK). I want to do some cluster analysis and I am aware that distance distortion might have an effect on the results. The precision I need is quite coarse (about 200m-500m). Do I need to take steps in the analysis or can I just calculate Euclidean distances?
-
Depends on the spatial extent of your data.– ErikCommented May 27, 2019 at 11:23
-
all over the UK– StrabonioCommented May 27, 2019 at 11:24
-
3If a national grid isn't accurate to 200-500m, it would be an epic failure.– VinceCommented May 27, 2019 at 11:56
-
1@Vince: I guess OP is asking whether he should take the ellipsoid into account or could go ahead calculating Euclidean distances.– ErikCommented May 27, 2019 at 12:05
-
If it's worth asking the question, it's probably worth the effort to do the math. Use a subset of the points which Euclidean distance says are farthest apart, deproject, and then calculate distance on the spheroid. I'd expect that the difference would be on the order of a meter in such a small country.– VinceCommented May 27, 2019 at 12:28
|
Show 1 more comment
1 Answer
Using NUTS-3 centroids with the Near (Analysis) tool, the geodetic distance between the Shetland Islands and Jersey was calculated to be 1191878.03 meters. Using a Euclidean distance after projecting to British National Grid (with the default datum transformation), those two points are 1191362.07 meters apart (a delta of 515.96 meters, or 0.043 percent).
The distance between the centroids of Westminster and Edinburgh is 532,119.78 (geodetic) and 531,949.88 (Euclidean) with a difference of 169.90m (0.032%), and between Westminster and Belfast is 519,194.12 (geodetic) and 519,162.44 (Euclidean), with a difference of 31.69 meters (0.006%). Using closer points, Sheffield and Westminster are 229,965.15m (geodetic) and 229,897.19m (Euclidean) apart, with a difference of 67.96m (0.030%)
So, with just four samples, it seems the intrinsic error over the British National Grid is not more than 0.05%, which hovers close to your desired precision. You'll probably want to run the numbers both ways in your analysis, then make your choice from there.
-
I'm guessing that for cluster analysis you won't have any points that far apart Commented May 27, 2019 at 17:31