Arctic ice "health index"

[SimonF92]

Hi

I havent been around much for a few years. I work in artificial intelligence. Without committing to much personally, a lot of AI is data processing and a lot of the data processing is working with images. I have been pulling Arctic satellite data recently and been thinking about a good summary metric to describe ice "health". People have done this before in this forum. I remember many years ago someone came up with a SMOS beige pixels algorithm that was really nice. I wonder if i can do something along those lines.

Im looking to create a summary metric. This can be regional but it can also be whole-Arctic.

I can write (and share) the code. But im looking for some expert opinions on what ice "health" would look like when viewed from above. I'm considering concatenating these features in some way:

1) area/ extent
2) normalised reflective wavelength
3) floe size
4) distance between floes
5) floe motility & direction
6) change in any of the above
7) surface temperature
local albedo
9) recent cloud cover(?)

Im not an expert. Is there anything else people think should be in there?

Theres a whole field of AI called graph-convolution. I have been considering turning the floes into graphs but I think that might be overkill. Its also vulnerable to clouds. Thats another discussion but ill drop a hint to it here anyway.

[kassy]

I think coming up with a definition is the hard part.

The best way to do this is to come up with something which approaches a realistic 3D model of the Arctic ice but no idea how to get there. And how do you define ice health? If we knew the actual thickness at maximum SIA that might tell us something about the next season but that is then influenced by the pesky weather thing.

Health as a current state also does not tell us much so maybe get it to find a health which does predictions on at least Slaters time frame? That way there is at least something to verify.


 

[John_the_Younger]

I suspect ice age is a reasonable variable to add in determining ice health.  A floe with a pixel of old ice would be healthier than a similar floe with only 1st year ice as the old-ice pixels would be less salty.

Although the estimated ice thickness (and volume) are mostly just educated guesses, this may be another variable worth incorporating.

Accumulated freezing-degree-days may be another tool to use to help discern ice health.

I wonder if geography should be a factor.  Ice in Fram Strait will not last as long as ice half way between Ellesmere Island and the North Pole.

[Bruce Steele]

It seems meltponding and color of the ice are visuals that indicate health or the lack thereof.
That is lots of light blue ice and meltponds in early June portend potential for big melt seasons.

[SimonF92]

Thank you everyone for the input so far.

Kassy, agreed- even generating a label to see if this is useful is quite difficult. I will look at thickness also.

John, thank you. Ice age and thickness I should be able to pull from Bremen.

Bruce, thanks. Ice visual characteristics should be achievable- clouds can cause issues but im pulling .nc directly from Bremen where I can see they have already figured that out.

Progress incase people are interested in following along- im attacking fragmentation first.

[SimonF92]

I created some heuristics to describe fragmentation within the CAB.

NB these data only describe days in August.

Im not sure it shows much of a trend, ill plot the data more properly once I have cleaned the code a bit. 2016 was clearly the worst year for fragmentation in the CAB.

[uniquorn]

Looks interesting. How are you training?

[SimonF92]

Hey uniquorn, hope you have been doing well.

That's a really good question. I have basically complained that I don't think area/extent (less so volume) are good metrics - so what's the point in trying to predict them?

You have experience with segmentation style models like u-nets. The features I am engineering could be used as inputs to that kind of model (ice age masks maybe) in one of the smaller layers for inductive bias- but I'm not sure I see the point.

Using AI here could be like a hammer looking for a nail. At the moment I'm just interested in looking for patterns in the data.

Some self-supervised clustering could be cool though- to look at what years are similar. I'll look into that.

[uniquorn]

I hope that you and yours are doing well also.

The fragmentation analysis looks useful. I would add the LEADS component of SIC-LEADS and ASCAT (it's the best imo)
https://manati.star.nesdis.noaa.gov/ascat_images/ascatB_ice_image/msfa-NHe-a-2024251.sir.gif

missing files can often be found at https://www.ospo.noaa.gov/cData/Atmosphere/ASCAT/assets/20240806_ASCATice-MetOp-C.txt

[SimonF92]

That looks nice.

Bremen 3.125km data  suffers badly from clouds meaning I have to apply smoothing.

How are you pulling SIC-LEAD data? Manually? I cant find a good server I can write requests to.

August fragment mobility as a bribe (ice most mobile this year)

[SimonF92]

The findings as they currently are

JUNE, JULY, AUGUST 2013-2024

I will move onto area/extent, ice age and colouration.

Code used to make the above plot (it might not be great to use at the moment):
https://github.com/SimonFisher92/SeaIceHealth

[uniquorn]

SIC-LEADS files are only on ftp (I think) at ftp.awi.de/sea_ice/product/amsr2
leads only are wrapped up in the nh_SIC_LEAD_*.nc.gz files

Will take a look at the code over the next few days.

[SimonF92]

SIC-LEADS files are only on ftp (I think) at ftp.awi.de/sea_ice/product/amsr2
leads only are wrapped up in the nh_SIC_LEAD_*.nc.gz files

Will take a look at the code over the next few days.

Thanks uniquorn, im a bit of a weekend warrior on this codebase but i will try to incorporate SIC-LEADS this weekend

[Jim Hunt]

Bremen 3.125km data suffers badly from clouds

Could you classify "ice" versus "ice under cloud" versus "water under cloud" using the false colour version of the CIS weekly MODIS mosaic?

https://ice-glaces.ec.gc.ca/cgi-bin/getprod.pl?prodid=MODISCOM-F

[SimonF92]

Bremen 3.125km data suffers badly from clouds

Could you classify "ice" versus "ice under cloud" versus "water under cloud" using the false colour version of the CIS weekly MODIS mosaic?

https://ice-glaces.ec.gc.ca/cgi-bin/getprod.pl?prodid=MODISCOM-F

"Dont reinvent the wheel" is one of the best pieces of advice anyone ever gave me. While this is a great idea, I suspect others have thought of it before. Its very likely that some of the data we see day-to-day about sea ice has already had this done to it. I would rather track that down.

While we could train a model for that, it would require a lot of labelling. I already have a labelling tool written for some Scottish snow patch work- but it would require crowdsourcing volunteers to draw where the clouds are
(https://github.com/SimonFisher92/Scottish_Snow/blob/10-new_copernicus_api_trial/src/machine_learning/snow_mask_generator.ipynb).

That's a BIG task and i'd rather rule out that those masks don't already exist somewhere. (Perhaps SIC-LEADS have already had this done to them?).


You could approach it in a non-ML traditional image processing manner but I think that could be prone to error. In the Scottish snow stuff above we were using as many bands as we could get from Copernicus and it was still somewhat tricky telling cloud from snow with a pixel-wise random forest as a baseline model.

[SimonF92]

uniquorn that was a fantastic tip about the LEAD data, what an amazing resource and they have evidently been doing some heavy processing to get it in that state.

I am much more confident now.

This is what I have been thinking about. Does the condition of these leads, their spatial configuration and their attributes (such as individual size, velocity etc) mean much? I have set this graph up to represent these data (a graph is an unstructured data modality of nodes and edges).

How do the lead graphs appear, how do they change and does that mean anything wrt ice health?

[kassy]

The product shows where ice moves more so it shows a degree of plasticity vs rigidity. Of course this all interacts with weather and winds.

[SimonF92]

The product shows where ice moves more so it shows a degree of plasticity vs rigidity. Of course this all interacts with weather and winds.

Hmm, good point about wind actually ill need to be careful I don't just come up with a way of measuring wind- that would be embarrassing. Thanks