tag:blogger.com,1999:blog-7729093380675162051.post4368475745478517076..comments2024-03-16T02:27:38.423+11:00Comments on moyhu: Making an even SST mesh on the globe.Nick Stokeshttp://www.blogger.com/profile/06377413236983002873noreply@blogger.comBlogger5125tag:blogger.com,1999:blog-7729093380675162051.post-65469629237121992812017-03-16T19:25:18.425+11:002017-03-16T19:25:18.425+11:00Olof,
"Nick, do you mean culling of data poin...Olof,<br /><i>"Nick, do you mean culling of data points literally"</i><br />Yes. I tried originally replacing the remaining points by local averages including those with culled, but found it didn't make much difference, and was complicated with land around. SST varies very smoothly; it is itself "optimal interpolation". Culling potentially adds noise, but there is so little compared with land.<br /><br />I did this partly thinking about GHCN V4. With rational culling, one might well end up with V3 :)<br />Nick Stokeshttps://www.blogger.com/profile/06377413236983002873noreply@blogger.comtag:blogger.com,1999:blog-7729093380675162051.post-84546503228534392522017-03-16T18:37:28.483+11:002017-03-16T18:37:28.483+11:00Nick, do you mean culling of data points literally...Nick, do you mean culling of data points literally, or are you just averaging them to reduce their numbers? The first alternative would lead to a larger loss of info and more noise.<br /><br />Implementing GHCNv4 would give a new redundancy problem. Do you think it is possible to use rural stations only, to get the numbers down? Giss effectively only use rural stations for long-term trends, due to their UHI correction...Olof Rhttps://www.blogger.com/profile/18244733455655978307noreply@blogger.comtag:blogger.com,1999:blog-7729093380675162051.post-78751691785034519112017-03-16T11:30:11.408+11:002017-03-16T11:30:11.408+11:00With a bit of work I bet we could improve the data...With a bit of work I bet we could improve the data pipeline a bit to get a nice interpolation that's not dependent on data density and handles the shores better (for some version of better). A constrained Delaunay, rather than the Delaunay triangulation you're computing, would let the shores be represented as precisely as you want. And we could add points to break down big triangles as well (or at least to limit how fast element size shrinks, which is associated with interpolation error).<br /><br />But it looks like mesh error leads to single-digit percent differences in anomalies. So it's likely not worth it.numerobisnoreply@blogger.comtag:blogger.com,1999:blog-7729093380675162051.post-8718609160843223932017-03-16T05:40:35.313+11:002017-03-16T05:40:35.313+11:00Numerobis,
You're right that there isn't t...Numerobis,<br />You're right that there isn't that much data (though meshing every month, when I have to, takes a while). The motive for culling is to get roughly comparable density between sea and land. Otherwise in the boundary regions there is a bias - the land can have no influence on the densely meshed sea, while sea data will fill in wherever there is a gap in land data, as there often is. That still happens, but not so much.<br /><br />Putting points freely requires interpolation of the SST data. Generally that's fine; SST is fairly smooth, and 2° data gives a good basis. It gets tricky near land. It would have been tricky for ERSST too; I hope they know more about it than I do. However, I don't think freely moving points would get a much better mesh than this method.<br /><br />On the general meshing procedure, I use a convex hull algorithm. So in culling, I try to ensure that a reasonable mesh could be found, but I rely on the hull process to find it.<br />Nick Stokeshttps://www.blogger.com/profile/06377413236983002873noreply@blogger.comtag:blogger.com,1999:blog-7729093380675162051.post-18193271896421577082017-03-16T02:27:15.563+11:002017-03-16T02:27:15.563+11:00I don't fully understand your meshing procedur...I don't fully understand your meshing procedure.<br /><br />Two questions: (1) why subsample rather than just using all the data, it doesn't sound like that much; (2) why limit yourself to using the 2-degree points from your input data rather than putting points on the globe freely, in a regular mesh -- including up in the Arctic.numerobisnoreply@blogger.com