DaniJonesOcean
commented
2 years ago From Erin Thomas:
I think SOM, like the GMM method you use, is probably susceptible to the same problem: If no data exists, it is unable to determine if there is a unique structure to separate from the others. I think your data set has pretty good spatial coverage, so I would be fairly confident that the classifications resulting from SOM would capture most of the profile ‘types’. That said, if the regions with missing data are unique from the rest of the existing data, SOM will fail to define a classification for this (can’t classify what you don’t know). Based on the profiles you show, I think SOM would find some interesting characterizations. SOM typically are defined with more than 4-5 classifications (I used 36 classification in my SOM space), so it might be interesting to see what structure appear/grouped together if you sort the data by temperature or salinity individually….
Inspired by talk from Erin Thomas at OSM22, Norwegian Meteorological Institute (erinet@met.no) [has now moved to LANL], "Self-organizing mapping to validate model temperature profiles"
Can neural networks extract useful information in regions of limited observations (e.g. the Weddell Gyre)?
Can we apply this to the SO-CHIC dataset? Could be exciting