AIdeas

[uniquorn]

ai4eo downloads are working now and we have a new test_upload file. It throws new runtime errors for me but may be related to our method so will do a default run to check.

Meanwhile, first results from curriculum learning, perhaps not behaving as expected?

[A-Team]

Maybe. A shame to revert to pixel counting when we already have the lat/lons.

If all the twists and turns of buoy trajectories and isochrons were approximated/replaced by great circle segments based on lat lon end points, a canned library could step in to replace areal pixel counting.

Getting away from those and factoring in the elliptical surface of the earth leaves the world of geometry for that of applied numerical analysis. The mercator projection (tangent plane) would not introduce significant error here and no doubt the navy has some function converting its areas to actual sea surface areas.

However the 'grow selection by 1' tool in Gimp is good for adding border edges and the essential bounding of error. In some ways, the numerical is of less interest than the overlay which means the underlying satellite resolution is limiting.

At the end of the day, we know very little about ridging, narrow leads and thickness so it's best not to get pulled in to too many significant digits by the  overly generous accuracy of buoy positions and times.

"If it ain't got that swing, it don't mean a thing" -- Duke Ellington

Wow, 88 teams, great to see the interest. Who knew! 96 days to go, with today the first since November with all the download ducks in a row (?)

Maybe a timeout should be called until the critically important advances in AMSR2 super-resolution can be disseminated(?) However they are focusing on  areas far to the south of the Arctic Ocean where the ice will be highly seasonal and changing hourly.

It is almost like they are looking to track icebergs rather than ice. Which is what the commercial clients wanted in the 'stakeholder's' survey above: ice/ no ice.

Iceberg Detection in Dual-Polarized C-Band SAR Imagery
https://ieeexplore.ieee.org/abstract/document/9406184  April 2021

Field tracking (GPS) of ten icebergs in eastern Baffin Bay
https://tinyurl.com/2zdjsmyf  2017

stokholm 1 hour ago Dear participants, after a long back-and-forth discussion with our data hosting providers, we are happy to announce that the dataset again (all train and test versions) is available for download. To explain the problem in brief; there was a malfunction with a server where a large portion of the dataset was stored. Simultaneously, the person responsible for the server maintenance was on long-term sick leave making it very difficult to pinpoint the exact causation. I can assure you that the service provider has been working very hard to remedy the situation, which we highly appreciate.  On behalf of the competition team, I sincerely apologise for the inconvenience and frustration caused by this. Please let us know if you experience otherwise.
In other news, there was a small update to the test_upload.py file in the github repository. Thank you to user "ct" for pointing this out.  I have forwarded the issue to the AI4EO platform team.

[A-Team]

It seems the ice competition does not break out blackbox vs explainable AI entries, though it sounds like during the course of internal experimentation with parameters and approaches even blackboxers would have a very good idea of what the key elements were.

I'd say the scientific community had very little interest in blackboxes even if they do slightly better. But if google can be believed, there is a tool to bridge the gap.

Climate Signals And Explainable AI
https://zacklabe.com/climate-signals-and-explainable-ai/

By applying nonlinear neural networks and explainable machine learning methods (e.g., layer-wise relevance propagation or integrated gradients), we aim to disentangle forced climate patterns from internal variability in observations and large ensembles.

In particular, we are interested in using these methods to detect biases/differences in their simulation of compound extreme events, internal variability, and forced trends in fully-coupled climate models. Explainable AI methods can be used as another tool to understand physical mechanisms in the climate system.

By applying nonlinear neural networks and explainable machine learning methods (e.g., layer-wise relevance propagation or integrated gradients), we aim to disentangle forced climate patterns from internal variability in observations and large ensembles.

In particular, we are interested in using these methods to detect biases/differences in their simulation of compound extreme events, internal variability, and forced trends in fully-coupled climate models.

Explainable AI methods can be used as another tool to understand physical mechanisms in the climate system.

Comparing Climate Models And Observations In The Arctic
 Arctic Amplification ... Surface Temperatures
Predicting Temporary Slowdowns In Decadal Warming
Disentangling Aerosols And Greenhouse Gases

[SimonF92]

ai4eo downloads are working now and we have a new test_upload file. It throws new runtime errors for me but may be related to our method so will do a default run to check.

Meanwhile, first results from curriculum learning, perhaps not behaving as expected?

uniquorn, i saw your message come through on the slack via my emails but I couldnt respond. The validation set will fail on my branch, because i have tweaked (reduced) the input channels, the tensors will be the wrong shape as a result (ie they dont line up between the shapes the model was trained on and the shapes it sees in validation).

Ie we have commented out ('#') a few of the inputs, I expect the error will align exactly on one dimension with how many we reduced it by. I think there is more explaination on my part in .utils where you might be able to fix it. Its on my list to code up a better way of removing channels, but it doesnt address the problem that validation holdouts will always expect them all. One way to  hack around it is to set those channels to NaNs (at least thinking about it).

I will fix that tomorrow on my branch, a way to sanity check would be to see if that dataset works on your own branch.

A-Team, cool stuff, share the sentiments about black boxes, particularly research setting where incidental findings often lead to cool things

[uniquorn]

Gone back a few steps using mostly default code while I download the data. @sf92 Understood about the torch shape. 11b has good learning curve but never gets a better score.

<>In other words, should an AI offering be penalized because it differs from manual? It might very well be better.<>

Based on what I have seen so far, this, more than any other problems, may stop me investing too much more time in the project. Still curious about the methods though.

[uniquorn]

It's possible to improve the ai4eo score by weighting the parameters and data choices towards the blocky ice charts and away from the detailed SAR's but I prefer the more realistic granular results. Will give it a bit longer to see how it develops.


Epoch 10 training score:        submission score
SIC r2_metric: 74.317%                59.25%
SOD f1_metric: 88.826%               74.63%
FLOE f1_metric: 81.501%              67.39%
Combined score: 81.557%            67.03%
Note that the submission test images have some very grainy SAR images.

[A-Team]

may stop me investing too much more time in the project.

It's worthwhile. Without good AI skills, can't stay in the game. Or even understand what the game is any more.

submission test images have some very grainy SAR images.

That satellite has been a huge disappointment across the board. It just does not give good visually interpretable resolution. The most interesting thing about it is characterizing what is wrong with it.

improve the ai4eo score by weighting the parameters and data choices towards the blocky ice charts [while degrading actual prediction accuracy

That is reminiscent of weighting commercial weather prediction towards scary weather which enhances viewership (and ad prices, the seamy side of AI). See drought page post #514.

Big floe, small floe, cake ice? From the Polarstern days, I am vaguely recalling that the risk to icebreakers (and commercial shipping) was not sea ice but pieces broken off from glaciers or landfast, bergy bits, growlers or some such.

[uniquorn]

Still tinkering with variables on Stokholm's updated code and data before returning to curriculum learning. 14c, the latest effort, shows some promise after 29 epochs. Might be able to squeeze a bit more out of it with a longer run. Detects open water quite well in dispersed ice, some definition between old ice and thick FYI. Very poor at low SIC.
Could do with some competition now as a benchmark.

Epoch 26 score:                         submission score:
SIC r2_metric: 73.663%              67.134%
SOD f1_metric: 83.498%             73.157%
FLOE f1_metric: 78.963%            71.116%
Combined score: 78.657%          70.34%

[uniquorn]

Didn't follow up on a longer run of 14c, instead choosing to look for better results for low SIC. This morning's epoch scores were disappointing at 65.6% but submitted it anyway. Surprised to get a new high score of 74%

Much better gradation of SIC and FLOE. Might need a different scoring method.

Epoch 22 score:                               submission:
SIC r2_metric: 69.164%                  77.144
SOD f1_metric: 66.85%                   73.113
FLOE f1_metric: 55.541%                69.495
Combined score: 65.514%              74.002

[uniquorn]

<>
  Up to now we choose not to post code on their computer, though SimonF92 may benefit from its superior format. <>

Scratch that, the best default set up has 7.5 cpu, 1 gpu and 30GB RAM but it's a fair bit slower than my home PC when running almost the same code. Kernel restarted 70% through the first epoch, probably due to lack of memory. Could lower the mem demand if it was faster but there is a soft limit of 40hrs usage so not worth the hassle.

[uniquorn]

A very small improvement of 0.18% with the last run.

Very lonely on the leaderboard so far.
https://platform.ai4eo.eu/auto-ice/leaderboard

Epoch 23 score:                     submission score
SIC r2_metric: 72.19%           74.994
SOD f1_metric: 71.032%        74.165
FLOE f1_metric: 64.307%       72.591
Combined score: 70.15%       74.182

Starting to get more detail. Have some ideas for a pan arctic project using v110, worldview brightness temp and cs2/smos thickness.

[A-Team]

Have ideas for a pan arctic project using v110 AMSR2, worldview brightness temp and cs2/smos thickness.

Same software repurposed to display winter thermal gradient through ice and thus ice growth rate, a continuous thermistor that could be calibrated (or spot-checked for accuracy) with existing sparse buoys? Disruptive!

lonely on the leaderboard so far

Teams with no previous ice experience may not have the necessary commonsense. Or maybe like an eBay auction where a silent bidder shows up at the last second.

[uniquorn]

Maybe add this

Overview

Arctic Sea and Ice surface temperature product based upon observations from the Metop_A AVHRR instrument. The product is a daily interpolated field with a 0.05 degrees resolution, and covers surface temperatures in the ocean, the sea ice and the marginal ice zone.

DOI (product):
https://doi.org/10.48670/moi-00130

[A-Team]

daily interpolated field with a 0.05 degrees resolution

1º = 111 km  (60 nautical miles)
0.01º = 1.11 km (2 decimals, km accuracy)
0.05º = 5.55 km = 5550 m

Fairly large pixels relative to intrinsic size of ice thermal features such as long but narrow leads or inter-floe distances but still providing a good smoothed overview on temperatures?

Wondering about synchronicity of various satellite sources if ice is moving 10-20 km/day and if there is any benefit to correcting for motion (as provided by buoys or more easily OsiSaf).

Must not be too affected by clouds or they couldn't even offer it.

Arctic OceanLat 58° to 90°Lon -180° to 180°
Since 4 May 2019
Blue markets
blue (physical) white (sea ice) and green (biogeochemical) ocean state
WGS 84 (EPSG 4326)
Updated daily 13:00 UTC
NetCDF-3
DMI (Denmark) Spotted an error?  Send us a note through the chat!

[uniquorn]

Collating the data will be a big ask, no idea how to correct for motion.

Lots of low SIC on the latest run. There may be a problem with submissions, 2 of ours are pending now, last nights has been pending for 13hrs. Maybe get fixed on a weekday.

[SimonF92]

Update from me.

Uniquorn and I have gone off on different tangents. Initially we forked a repo written by someone from the project. The repo is definitely well written, but often if you use someone elses code, it can be quite hard to make meaningful modifications.

So while uniquorn is continuing to explore the hyperparameter-space of the task on the forked repo, I am writing a model from scratch and heavily annotating the code.

Its slow going and unglamorous but I am definitely still invested on getting something out of this. I think in time it will be beneficial to have our own codebase, but its going to take me a while.

Im continually mulling over, and asking questions about, coding-in the temporal component of the ice data. I dont think a U-net is optimal because it makes no use of the date (at least explicitly):

U-Time: A Fully Convolutional Network for Time
Series Segmentation Applied to Sleep Staging
https://proceedings.neurips.cc/paper/2019/file/57bafb2c2dfeefba931bb03a835b1fa9-Paper.pdf

[uniquorn]

Im continually mulling over, and asking questions about, coding-in the temporal component of the ice data. I dont think a U-net is optimal because it makes no use of the date (at least explicitly):

U-Time: A Fully Convolutional Network for Time
Series Segmentation Applied to Sleep Staging
https://proceedings.neurips.cc/paper/2019/file/57bafb2c2dfeefba931bb03a835b1fa9-Paper.pdf

Looks interesting though that must be a long run. Smaller data maybe.

Training of U-Time was stopped after 150 consecutive epochs of no validation loss improvement

[uniquorn]

Some improvement with the tweaking on the latest run. Quite a long run but optimised to be a little quicker.

Epoch 40 score:                        submission score:
SIC r2_metric: 77.769%           76.70
SOD f1_metric: 69.956%          76.08
FLOE f1_metric: 70.011%         73.39
Combined score: 73.092%       75.79

SIC starting to look quite realistic.

[uniquorn]

On run 29a above epoch 40 and 43 had very close local combined scores at 73.092% and 72.983%, a difference of 0.109% yet the submission scores differed by 1.44%.

Seems a bit hit and miss. 15b up thread had a massive score difference of 7.98%.

[SimonF92]

uniquorn, not at my own machine at the mo, and I may be misinterpreting this,

but I hope you are keeping in mind that if you continually submit models to their holdout set online they may start to suspect you are (unintentionally) overfitting to the test set and would have some things to say about that most likely

[uniquorn]

Run 29a took 16 hours for 53 epochs, based on the evidence it's quite possible that a slightly lower local score may give a higher submission score. There's no other way to find out than submitting. It's not like I'm keeping anything secret, apart from the variables and data choices.

I like Warewolf's approach, posting a negative score as a starting point.
https://platform.ai4eo.eu/auto-ice/leaderboard
I think that's only possible if you've already edited the default code.

We could make a new account and only submit the best models occasionally. Like asif profiles

[uniquorn]

This image generated by run 31a has a lot of good spatial detail despite the wave interference. SIC is as expected. Interesting that the tendrils are detected as thick FYI and big floe. Young ice in march?  You could be right about needing a date method.

inf31e51 is the second highest local score

[uniquorn]

Trying a new tack. Bin filling. Submissions process crashed again.

[uniquorn]

So it might be possible to run an AI on a series of these.
max full arctic res at 1km on wv is not bad, evaluate with sic-leads for cloud cover, chuck in sea/surface temp and cs2smos for good measure..

[uniquorn]

Some submissions have been processed and we have a new high score of 76.36% but that is only enough for second place. Rompetechos has 76.59% 

[A-Team]

Back on top! Looks like 'Floe' has the largest opportunity for improvement with ArcticTerns having 71.73 improving on leader 69.53 by 2.20 which would put eciaes just over 80 which is still nowhere near 100.

Maybe next time the contest can be turned around -- AI is likely far more accurate already -- and manually drawn ice charters can struggle to get out of the 70's.

Only 12 teams active out of 85 with 69 days to go. Some of the 73 missing teams may not show their cards until the last minute  though more likely they dropped out after an initial bout of enthusiasm.

https://platform.ai4eo.eu/auto-ice/leaderboard

[uniquorn]

The last run was longer again but each epoch a little faster. Here is an example of the top two local scores. We are looking for broadly similar SIC, SOD and FLOE. Widely differing results would suggest that the model is wrong.

[uniquorn]

Floe size and diff from the 2 highest scores on the latest run.

[A-Team]

It appears that the vast majority of teams will never submit anything and that the leadership has already hit the wall. Should the scoring system should even be providing one decimal point? Providing two implies a broader - and troubling - misunderstanding of the scientific process.

[uniquorn]

Might have gone as far as I can with the current method.

So it might be possible to run an AI on a series of these.
max full arctic res at 1km on wv is not bad, evaluate with sic-leads for cloud cover, chuck in sea/surface temp and cs2smos for good measure..

I don't know why I didn't include ascat in that list. jan1-feb14

[A-Team]

meant to include ascat in that list. jan1-feb14

Yes it has good information about the ice, fairly consistent display of features, little interference from weather, quite a bit of archival time depth and is straightforward to work with.

Note enhancement choices are a form of AI in themselves. The holy grail overall is putting together three distinct but possibly correlated sources of synchronous information into an RGB time series, with each making good use of its gray space so that the final product makes good use of its color space (ie is not too gray).

One of the proprietary Canadian radar products had this but we just saw a single image during the Polarstern year.

Nullschool had a textbook wind and pressure pattern the other day, winds just right for both Transatlantic Drift and Beaufort Gyre. However the former has been surging westward towards Nares this year and the latter won't persist long enough to make the turn north at the Chukchi, the wind is weaker there already in the frame.

Note the Lincoln and Wandel Seas appear green here whereas more of the old thick ice gold might have been expected. The turnover is quite high above Greenland and the ice is now longer stable as historically presented.

[oren]

In addition to the Ascat animation, I find A-Team's first gif highly informative. The orange regions are the thick ice (so it seems) and their export or drift to a safer location can determine the setup for the melting season.

[uniquorn]

https://library.wmo.int/doc_num.php?explnum_id=9270

SIGRID-3: A Vector Archive Format for Sea Ice Charts

1. Introduction
Through the International Ice Charting Working Group (IICWG), the world's ice centers
developed a vector format for archiving digital ice charts. The ice centers most actively involved
in this effort are the Arctic and Antarctic Research Institute, Russia (AARI), the Canadian Ice
Service (CIS), the Danish Meteorological Institute (DMI), and the U.S. National Ice Center (NIC).
This new archive format joins the current World Meteorological Organization (WMO) standards for
ice charts in the Global Digital Sea Ice Data Bank (GDSIDB). WMO ice chart archive formats are
the Sea Ice Grid (SIGRID) format developed in 1981 and formalized in 1989, and its successor
SIGRID-2. The vector format defined in this document, SIGRID-3, joins SIGRID and SIGRID-2 as
standard WMO formats.
SIGRID-3 is based on a format called “shapefile.” The shapefile format is an open vector
file format (see Appendix 2 for more information). In contrast to raster formats such as SIGRID-2,
where ice characteristics are represented on a grid, vector formats represent features (such as
areas of ice outlined on a chart) as a series of vertices that define the outline of the feature in
space. An associated list of attributes (such as the concentration, stage of development, and
form of ice within) characterizes ice within the outlined area.
Storing ice chart data in vector format rather than raster format has advantages. The
vector file preserves all of the information in the original chart, and charts can be re-projected or
re-scaled without loss of information. It is also possible to convert a vector product to raster if
necessary. These qualities make the vector format attractive to the researchers who are the main
users of the GDSIDB. In addition, charts in SIGRID-3 format will be easy for ice centers to
produce using many of the current production systems that employ Geographical Information
Systems (GIS). Shapefiles can be produced without commercial GIS software but this requires
the development of custom software

[A-Team]

vector GIS for ice, nightmare shapefile format
https://en.wikipedia.org/wiki/Shapefile

This is a really bad idea. Stick with raster.

This is ESRI and arcGIS, huge learning curve, constant crashes. Ditto Google Earth.

Vectors are great for city features: urban growth boundary, underground utilities, sewer lines, property lines, power lines, street names, roads, intersections, building footprints, need for scalability, tons of text.

Vectors are terrible for natural features. Imaging data comes intrinsically in pixel arrays, not polygonal blocks. Raster has fantastic rescaling but is little needed in Arctic imagery as the resolution is almost always hurting to begin with. All computer screens including CRT are raster.

Sometimes a vector overlay on raster is useful. Gimp offers that as Paths Tool but it is not easy to use.

I can share a vector Arctic coastline file if you would like a polygon with ten thousand vertices.

To me, this is the heart of the whole problem with Ice Charts. A consultant told them ages ago to dumb it down to polygons. Someone sits at a computer looking at satellite picture and clicks over it with their mouse and saves out a rough polygon that greatly simplifies the data.

They went for vector and now the AI contest which is inherently raster, has to be dumbed down to match their blocks and lines.

[uniquorn]

Comparison of the top 5 local scores and submission scores of run 38a.
It's still like LOTO

[uniquorn]

They went for vector and now the AI contest which is inherently raster, has to be dumbed down to match their blocks and lines.

Latest run didn't improve score. I have some ideas how to dumb down though. Is it worth it?

SARs
latest best score (may not be the most accurate)
expert ice chart (no disrespect)
WMO egg codes  https://ocean.dmi.dk/polarview/egg_codes/egg_codes.php

[A-Team]

I have some ideas how to dumb AI raster down to vector though. Is it worth it?

Sure. As an exercise. Adobe Photoshop to Adobe Illustrator. Just picture yourself clicking around on a photo tiling up basic ice chart polygons.

Can be done in Gimp without the Path Tool by making a transparent new blank layer, activating the "Free Select Tool" next to the oval tool, making a simple polygon over the satellite layer, and filling. Then make a new blank layer, restrict to the complement of the previous polygon, and make a second polygon.... The main thing is partitioning: when you are done, all the original pixels are assigned to a unique polygon.

To do all at once, tile up all the images, simplify with color wand and 'Grow Selection' to greatly reduce polygons. I recall both ImageJ and Gimp can vectorify from that, as can various online tools because this is done a lot in animating.

[A-Team]

Now I'm wondering if you would be better off actually vectorizing your product, simplifying that, and re-rastering to blocks before submitting.

For example, a big reason why urban planners like ArcGis is that they can quickly add up linear miles of underground sewer lines or calculate the area of city rooftops available for solar, and so on.

In your situation, the AI resolution being too good translates into many polygons whose areas are too small (stray pixels inside an otherwise solid block).

If a cutoff were applied to polygon areas, that would have the effect of cleaning up blocks after re-rasterization.

If a boundary is too ragged (too many vectors), small concavities could be removed.

These are likely editing click options within ArcGis as shapefiles are often too complex or even contradictory (polygon crossing itself, not closing, having zero area).

Stray thought: that was interesting yesterday, the Next Big Thing (ChatGPT unicorn) being exposed as a hype. From word-smithing far too much myself, I was sure their early responses had been written or polished by English literature majors (once the dominant product of our universities).

With those folks largely being under-employed work from homers and a billion dollars available for cash burn, it was surprisingly affordable. They could not respond in real time to internet users so how was it possible to come up in real time with these beautifully crafted answers?

Well, it is the same with passwords: not that many distinct answers needed. After 125 million eight character passwords were stolen from Adobe, it quickly emerged that a few dozen were used hundreds of thousands of times, eg 12345678. It is the same with internet questions as Bing and Google well know.

[uniquorn]

Thanks for the ideas, concentrating on nrt for now.

Explainable Machine Learning for Scientific Insights and Discoveries
Ribana Roscher; Bastian Bohn; Marco F. Duarte; Jochen Garcke
24 February 2020    https://ieeexplore.ieee.org/document/9007737

Abstract:
Machine learning methods have been remarkably successful for a wide range of application areas in the extraction of essential information from data. An exciting and relatively recent development is the uptake of machine learning in the natural sciences, where the major goal is to obtain novel scientific insights and discoveries from observational or simulated data. A prerequisite for obtaining a scientific outcome is domain knowledge, which is needed to gain explainability, but also to enhance scientific consistency. In this article, we review explainable machine learning in view of applications in the natural sciences and discuss three core elements that we identified as relevant in this context: transparency, interpretability, and explainability. With respect to these core elements, we provide a survey of recent scientific works that incorporate machine learning and the way that explainable machine learning is used in combination with domain knowledge from the application areas.

The basic ML chain learns a model from given data and provides an output. Given the model and input-output relations, a scientific outcome can be derived by explaining the output results utilizing domain knowledge. A transparent and interpretable model can be explained using domain knowledge leading to scientific outcomes. Additionally, the incorporation of domain knowledge can promote scientifically consistent solutions

---------------------

Three teams with exactly the same scores today. Band split due to musical differences perhaps.
https://platform.ai4eo.eu/auto-ice/leaderboard

[A-Team]

Seeing is believing(?) It appears that UW_VIP has shared its code to jumpstart two other teams. If one of these later takes the lead (from University of Waterloo's Systems Design Engineering), the consortium can then replicate so that they take prizes 1-3.

[uniquorn]

2 days in the life of a very slow AI.
day58 run for SIC.

[uniquorn]

New high score of exactly 80.0 puts us temporarily in 10th place. This run was long (for me) with a very small patch size of 96px allowing a larger batch size of 17, possibly could be more but using less memory causes less interference with other programs. cpu usage tends to be high during the training stage, memory usage is high during validation.
Using a smaller patch size might enable more detail in the charts with the possible downside of amplifying background noise.

1. Two epochs had a very similar combined score next to each other. I thought it might be interesting to look at the resulting ice charts and their difference. There is reasonable agreement on open water.

epoch combi   SIC            SOD        FLOE
333   77.49    78.626    79.115    71.967
334   77.488   79.519   77.683   73.037

333sub  79.138   84.432   79.127   68.571
334sub  78.769   85.328   77.242   68.703

For both the SIC score is much higher, FLOE much lower.


2. Example of patch sizes on an S1A image.

[uniquorn]

That was fun. Time to leave AI to the research students and get some peas in the ground.
A messy chart of the end of the last run. Thunderstorm and power cut stopped play.

[A-Team]

peas in the ground

Very respectable effort!

I do wonder about the computer power needed here and whether the contest results in any better understanding of the ice.

I am quite impressed with this Crissy scoring this high on their first submission. Could be a winner (though current high 85 is a long ways off).

[uniquorn]

 Sam Altman: OpenAI CEO on GPT-4, ChatGPT, and the Future of AI | Lex Fridman Podcast #367

[vox_mundi]

New AI System Can Map Giant Icebergs From Satellite Images 10,000 Times Faster Than Humans
https://phys.org/news/2023-11-ai-giant-icebergs-satellite-images.html

Scientists have trained an artificial intelligence (AI) system to accurately map—in one-hundredth of a second—the surface area and outline of giant icebergs captured on satellite images. The paper, titled "Mapping the extent of giant Antarctic icebergs with Deep Learning," is published in The Cryosphere.



It is a major advance on existing automated systems, which struggle to distinguish icebergs from other features in the image. Manual—or human—interpretation of the image is more accurate, but it can take several minutes to delineate the outline of a single iceberg. If that has to be repeated numerous times, the process quickly becomes time-consuming and laborious.

"Using the new AI system overcomes some of the problems with existing automated approaches, which can struggle to distinguish between icebergs and other ice floating on the sea or even a nearby coastline which are present in the same image."

Dr. Braakmann-Folgmann and her colleagues used an algorithm called U-net—a type of neural network—to "train" a computer to accurately map the outline of icebergs from images taken by Sentinel-1 satellites operated by the European Space Agency.

As part of the study, the effectiveness of the U-net algorithm was compared to two other state-of-the-art algorithms used to map icebergs. They are known as k-means and Otsu. The algorithms were programmed to identify the biggest iceberg in a series of satellite images.



The system has been tested on satellite images of seven icebergs, which were all between the size of the city of Bern—54 km2; and Hong Kong—1,052 km2. For each of these icebergs, up to 46 images were used that covered all seasons from 2014-2020.

Over a series of tests, U-net outperformed the other two algorithms and was more effective in delineating the outline of an iceberg in images taken when environmental conditions were challenging, such as the image capturing a lot of ice structures.

On average, the U-net algorithm showed only a 5% lower estimate of the area of an iceberg. In contrast, the k-means and Otsu algorithms returned—on average—figures for iceberg area that were 150% to 170% too large, probably because the algorithms were including sea ice and even nearby coastline in the calculations.

In machine learning, the F1 score is an evaluation of how well an algorithm performs and ranges from 0 to 1, with values closer to one displaying more precision. U-net achieved an F1 score of 0.84. The two other algorithms both scored 0.62.

Mapping the extent of giant Antarctic icebergs with Deep Learning, The Cryosphere (2023).
https://tc.copernicus.org/articles/17/4675/2023/