I think the U shape is good. If you have time I'd like to see how it moves with time. I might be able to chart it in R if not.
![](https://forum.arctic-sea-ice.net/index.php?PHPSESSID=f9a60621cba78812785568c5c636331f&action=dlattach;topic=3013.0;attach=145079;image)
I experimented with changing the number of rows from 10
5 rows was worse
20 rows removed the large drops but the algo is struggling to detect the ice/snow boundary so ice goes up when snow goes down.
at 50 rows other problems crept in.
So far it is good for detecting snow+ice thickness
I'll play with this line in the meantime
#select the region with large stdev snow
dfU=dfU.where(dfU.stdev_delta_mean<-0.25)
If we had a large training dataset
rofl. This is probably the biggest it's ever gonna get.
adding the Tice/water, Tsnow/ice numbers etc to the csv might help debugging, maybe I can do that?!?