I started to work on MODIS-A SST (processed from L1 to L3m, saving SST(11 and 4), Vis/NIR/BTs reflectances (TOA/BOA)) for my area, the North Sea.
When I look at a random map, let say the 20th of April 2005 (AQUA_MODIS.20050420T121505.L2.SST.nc, https://oceancolor.gsfc.nasa.gov/cgi/br ... =D&prm=SST), I see values from -40 to 20°C for this image. I can understand that the -40s are associated to the clouds, but then is it still a "sea surface" algorithm ? I masked the clouds using cldice for my processed data (because I initially wanted the visible signal) so it is not a real issue, but I would like to know if this is normal.
I have some questions :
1) What in situ measurements have been used to train the sst4 / sst11 algorithms ? Temperature measurements above the sea ? At "1" meter depth ? Something else ?
2) How good should the algorithm be compared to the in situ data (the usual 1 m), what difference should I expect ? +/- 1°C ? 5 ?
3) Should I use the sst 4 or sst 11 algorithm for daylight comparison with my in situ sea surface temperatures ?
I have used SST, chlor_a, and other variables to pair in situ observations with remote-sensing observations to track dynamic changes in the boundaries of Longhurst provinces. This needs simultaneous SST and chlor_a, so night data are not useful, and I can live with quite large errors as the in situ data are sparse (so the matchups don't change much if noise is added to SST). SST and chlor_a from the same pass sometimes have very small overlap because chlor_a is more affected by sun glint and while SST is more affected by atmospheric path length. For other use cases, see:
Evaluation and selection of SST regression algorithms for JPSS VIIRS. April 2014. Journal of Geophysical Research Atmospheres 119(8). DOI:10.1002/2013JD020637
You are right, my good definition would be a distribution following the 1:1 line when we compare in situ and remote data, with something like +/- 3 times the % error associated to the in situ measurement. I would be happy with a SST algorithm returning +/- 1 °C compared to ground data for example.
While most chla algorithm publications have a matchup comparison, it is not something I see a lot in SST publications, like the one you referred to, this is the reason of my post, I always prefer visual comparison rather than metrics.
I do have some in situ values for my area (several thousands) paired with remote-sensing observations, but I had a +/- 10°C when I compared it to the SST11 / SST4, so I thought I was doing something wrong, which was the initial reason of my post.