murraycutforth
commented
10 months ago Okay I've taken this approach a step further, and computed the area of different pixel types over time, in each ROI. Hopefully showing the resulting plots will make it clear!
The snow cover looks sensible at a high level. It peaks every winter, and there is even the double peak visible in 2022/2023 which I remember well! But I still think this is all based on SCL masks which are very unreliable.
Next steps- add more ROIs for specific snow patches?
The sentinel data products are arranged in 100km^2 tiles, and I've downloaded all the 20m^2 SCL masks with cloud < 50% for tile T30VVJ, which corresponds to the Cairngorms. I'll show a bunch of examples here.
One issue with my approach so far, is that due to space limitations I haven't downloaded the image bands so I can't visually check these images.
These plots also show some mountain regions which I've manually marked out in my code.
1st April 2018
This mask looks quite good, with the land/sea/clouds/cloud shadows all being sensible. But I don't trust how the boundary of the snowy mountains is the light blue "thin cirrus" class. I think that given the spatial distribution it's got to be the edge of the snowpack..
27th June 2019
In this mask, almost all the land surface is misclassified as cloud shadow for some reason.
18th Jan 2021
This is an example gone very bad, loads of water classified over the land, and looks like an unreliable mixture of cloud and snow