TempLS is my global temperature anomaly calculator. It uses ERSST V4 for the sea surface component. ERSST unhelpfully records sea covered with ice as having the sea water freezing point of -1.8°C. Since this does not seem to be a good proxy for air temperature, I mark such points as missing data.
However, it seems -1.8 is not reliable - regions which are clearly frozen can report higher temperatures. Maybe there is an allowance for salinity variation. Anyway, to date I have been using ≤-1°C as the criterion for removal.
I now find that helps, but also doesn't work reliably. Even with the Arctic ice near peak, there seem to be a few deviant points reporting SST where there should be none. This creates a problem with my triangular mesh weighting in the Arctic. The idea is that land stations should be a better basis (than -1.8) for estimating nearby frozen seas. But these isolated spurious sea temperatures also become prominent in the mesh, and affect large areas, as if they were land.
So I am using a new criterion. For each SST location and month, I look at the record starting 1900, and count the occurrences of temperatures less than -1.5°C. If there are more than 10, I deem the location to be subject to intermittent freezing, and exclude it for that month throughout.
This will clearly exclude some valid data on the fringes of the ice. However, with some years frozen, others not, it is in any case hard to get an appropriate normal. In practice the decision about inclusion does not have dramatic effects for individual locations, since freezing leads to a zero anomaly. And because it is a lat/lon mesh, there is an artificially high node density anyway.
You may notice some small differences in the TempLS mesh results. The August average rose from 0.7°C to 0.703 °C.