Researchers create AI algorithm to detect cloud formations that lead to storms
Penn State
When forecasting weather, meteorologists use a number of models and data sources to track shapes and movements of clouds that could indicate severe storms. However, with increasingly expanding weather data sets and looming deadlines, it is nearly impossible for them to monitor all storm formations — especially smaller-scale ones — in real time.
Now, there is a computer model that can help forecasters recognize potential severe storms more quickly and accurately, thanks to a team of researchers at Penn State, AccuWeather, Inc., and the University of AlmerĆa in Spain. They have developed a framework based on machine learning linear classifiers — a kind of artificial intelligence — that detects rotational movements in clouds from satellite images that might have otherwise gone unnoticed. This AI solution ran on the Bridges supercomputer at the Pittsburgh Supercomputing Center.
Steve Wistar, senior forensic meteorologist at AccuWeather, said that having this tool to point his eye toward potentially threatening formations could help him to make a better forecast.
“The very best forecasting incorporates as much data as possible,” he said. “There’s so much to take in, as the atmosphere is infinitely complex. By using the models and the data we have [in front of us], we’re taking a snapshot of the most complete look of the atmosphere.”
In their study, the researchers worked with Wistar and other AccuWeather meteorologists to analyze more than 50,000 historical U.S. weather satellite images. In them, experts identified and labeled the shape and motion of “comma-shaped” clouds. These cloud patterns are strongly associated with cyclone formations, which can lead to severe weather events including hail, thunderstorms, high winds and blizzards.
Then, using computer vision and machine learning techniques, the researchers taught computers to automatically recognize and detect comma-shaped clouds in satellite images. The computers can then assist experts by pointing out in real time where, in an ocean of data, could they focus their attention in order to detect the onset of severe weather.
“Because the comma-shaped cloud is a visual indicator of severe weather events, our scheme can help meteorologists forecast such events,” said Rachel Zheng, a doctoral student in the College of Information Sciences and Technology at Penn State and the main researcher on the project.
The researchers found that their method can effectively detect comma-shaped clouds with 99 percent accuracy, at an average of 40 seconds per prediction. It was also able to predict 64 percent of severe weather events, outperforming other existing severe-weather detection methods.
“Our method can capture most human-labeled, comma-shaped clouds,” said Zheng. “Moreover, our method can detect some comma-shaped clouds before they are fully formed, and our detections are sometimes earlier than human eye recognition.”
“The calling of our business is to save lives and protect property,” added Wistar. “The more advanced notice to people that would be affected by a storm, the better we’re providing that service. We’re trying to get the best information out as early as possible.”
This project enhances earlier work between AccuWeather and a College of IST research group led by professor James Wang, who is the dissertation adviser of Zheng.
“We recognized when our collaboration began [with AccuWeather in 2010] that a significant challenge facing meteorologists and climatologists was in making sense of the vast and continually increasing amount of data generated by Earth observation satellites, radars and sensor networks,” said Wang. “It is essential to have computerized systems analyze and learn from the data so we can provide timely and proper interpretation of the data in time-sensitive applications such as severe-weather forecasting.”
He added, “This research is an early attempt to show feasibility of artificial intelligence-based interpretation of weather-related visual information to the research community. More research to integrate this approach with existing numerical weather-prediction models and other simulation models will likely make the weather forecast more accurate and useful to people.”
Concluded Wistar, “The benefit [of this research] is calling the attention of a very busy forecaster to something that may have otherwise been overlooked.”
###
In addition to Zheng, Wang and Wistar, the research team included Yukun Chen, doctoral student in the College of IST; Jianbo Ye, former doctoral student in the College of IST and current applied scientist at Amazon Lab 126; Jia Li, professor of statistics in Penn State’s Eberly College of Science; Jose Piedra-Fernandez, collaborating faculty member at the University of AlmerĆa, and Michael Steinberg, senior vice president at AccuWeather, Inc.
The researchers’ work was supported in part by the National Science Foundation, the Amazon AWS Cloud Credits for Research Program, and the NVIDIA Corporation’s GPU Grant Program, and was published in the June 6, 2019, issue of IEEE Transactions on Geoscience and Remote Sensing.
Just asking. Wouldn’t you be better off using real intelligence to forecast anything rather than the artificial stuff???
naw…this way we get a lot more false alarms…..that gives the impression global warming is making it worse
My first thought as well. They tout a 64% detection rate, “It was also able to predict 64 percent of severe weather events, outperforming other existing severe-weather detection methods.”
With a 64% detection rate does that equate to a 36% false negative solution? They missed identifying over a 3rd of the events.
How many false positives?
Have these researchers never heard the childhood story about the boy who cried wolf?
Yep, a bit like the joke that Economists successfully predicted 7 out of the last 5 recessions.
They sound plausible, and have done something interesting with past data where they get to choose all of the start and end points.
Now they just have to move into the real world and make some genuine, useful, predictions. Then we will see who is really cooking on gas.
That’s a nice way to say that they are giving up on the understanding of severe weather. The climate, on the other hand, is understood perfectly. Nothing more left to understand. The science is settled.
Here is the very first AI (artificial intelligence) that was developed for use by the public.
Reply hazy, try again
Joe Bastardi at WeatherBell Analytics is a meteorologist in the mold of Dr. Francis Davis who was a Physics Professor at Drexel University and was Forecaster at Philadelphia Station WPVI in the 60s & 70s.
Frank used historical data of the many weather systems that impacted SE Pennsylvania. He used to keep a diary and according to sources had a 92% forecasting accuracy.
Joe Bastardi has the advantage of the current models which quite honestly have questionable performance so Weatherbell has their own program developed by Joe B. & Joe D’Aleo and they combine their model with historical patterns to refine their accuracy.
Has anyone seen this nonsense— http://www.igsd.org/study-u-s-costal-communities-face-more-than-400-billion-in-seawall-costs-by-2040/
So every forecasting office gets its own supercomputer cluster?
“So every forecasting office gets its own supercomputer cluster?”
Not necessarily. It depends a lot on what usable forecasting algorithms look like. Maybe they can be simplified and the important parts run on anything that computes. And in any case, adding more computing cores is relatively cheap and getting cheaper.
AI is a clearly fad and we are probably (hopefully) near peak AI. A lot of AI stuff is probably nonsense, and some of it is potentially dangerous for a variety of reasons. You likely do not, for example, want an AI “algorithm” that no one understands steering your car down an expressway at 135kph. Nor, probably, do you want AI deciding who gets loans at what interest rates. At least I don’t want that.
But this seems fairly OK. The consequences of the algorithms being dead wrong sometimes or biased in strange ways are (probably) negligible. To be useful, it just has to be right a bit more often than humans in some situations. It is conceivable that the algorithm output — even if it turns out to be unreliable — can a meaningful input to human analysis.
We’ve already seen the limits of “self-driving” cars and it was not pretty. While humans don’t really affect weather like they do driving behavior they will have biases that get written into the algorithms. People write the algorithms. Just because a computer runs them doesn’t make the algorithms right, appropriate or better than human beings doing the thinking.
Sheri — There is at least one problem with AI along the lines you suggest. But AI doesn’t work quite as you suppose. What happens with AI is (sort of) “They” feed a generalized algorithm a bunch of weather data and the fact that severe weather occurred at certain times and places. The algorithm thinks for a while and says something along the line of “I see that 14 of 23 severe events were accompanied by X radar pattern and only 4 times did I see X without severe weather.
That actually can recognize correlations that people miss. Trouble is that it may embrace correlations that people would instantly reject. e.g. Severe weather only occurs on Tuesdays and Saturdays.
The known potential problem isn’t the algorithms. It’s that sometimes the training data is unintentionally biased. For example, a facial recognition algorithm trained on pictures of firemen in Iowa is likely to be poorer at recognizing orientals and Pacific Islanders than the same algorithm trained on pictures of the Honolulu FD.
you donāt understand machine learning AI. People donāt write the algorithms. The machines writes its own rules as it learns from thousands of simulations.
You donāt understand, ā¦. machine learning is a science fiction claim. Machines (computers) are simply great manipulators/processors of data/info based solely upon their āoperationalā programs (software/firmware). Software/firmware that was created by biological intelligence, ā¦.. a horse of a different color.
A prime example of the, per se, Artificial Intelligence (AI) is an āunabridged dictionaryā.
Everything you need to know is contained therein said ādictionaryā ā¦ā¦ and all one has to do is āselect the word(s) they needā ā¦ā¦. and then place them in the correct sequence.
The machines optimize the parameters of the algorithms written for them. AI is not a creative process.
@ n.n
If ā¦.. āAI is not a creative processā ā¦ā¦ then it should not be referred to as an āintelligentā entity.
Call it AS (artificial simulator), ā¦.. or AR (artificial reproducer), ā¦ā¦ or ABS (artificial BS), ā¦ā¦ but not intelligent or intelligence.
Typically AI is a sort of state machine. You initialize a program with multiple algorithms, in this case, to detect variations in clouds. Another program simply presents the Weather Machine with the photos. The Weather Machine analyzes the picture and gives it some kind of score- at the least its “Storm” vs “No Storm”. The Operating Program scores the Weather Machine according to whether or not actual storms were detected. If they were. those nodes in the program that predicted “Correct Answer”get modfied(rewarded) slightly to have a stronger response.
Doing this at hundreds or thousands of pictures a second the Weather Machine will start getting more and more correct hits. Eventually, (in a few hours) it will be able to identify most of the pictures that would lead to storm development.
Marvin Garder, writing in the “Real” Scientific American had a column where he described how to build a computer to play Tic-Tac-Toe perfectly- using 9 match boxes(I think, it’s been a while) and two colors of beads.
After a player marks a square he puts his color bead into the matching box. At the end of the game the player who won takes out the losing color from every square played. I think that is the right algorithm, but I’m sure it was more complicated.
After some dozens of “games” the computer could stalemate every game.
2 years ago there was an AI chess computer program called AlphaZero based on machine learning. It played a match against the top computer chess program called Stockfish which is a freely available public program. However the Alpha Zero people turned off the opening book of Stockfish. They claimed that AlphaZero didnt have an opening book so that it was unfair for Stockfish to use an opening book against AlphaZero. However reality is that Alpha Zero created its own opening book in the 100’s of thousands of games that it played against itself in training. So if the results of those games were taken out of Alpha Zero’s database, it wouldnt have any playing strength at all. So in effect it was indeed playing with its own opening book which at GM level and above is a huge advantage. No wonder AlphaZero won that match easily, which the Alpha Zero people claimed that it proved that the AI method was more successful than the traditional method of human tinkering. IT PROVED NO SUCH THING BECAUSE OF THE OPENING BOOK ADVANTAGE.
Alan T, ā¦. if one used two (2) identical Chess playing computer programs to compete against one another ā¦.. how in the world would they, per se, ālearnā anything, ā¦.. given the fact that they would both react the same to each otherās āmovesā?
‘…limits of āself-drivingā cars…’: Have you ever wonder what will happen when there are X number of āself-drivingā cars all trying to get to 500,000 (drivers) work in the morning. Which car will speed up or slow down to let someone in the center lane to move inside lane to exit. We might as well go back to single lane highways. Think of it as the Computer run choo – choo train with your personal compartment.
Actually, fully autonomous cars will likely tend to be overly cautious because the manufacturers don’t want to be sued. They will also (probably) talk to each other and negotiate who gets to occupy what space when. Conceptually, that could be beneficial to traffic flow. However, doing that without major costly (or lethal) mistakes is going to be a huge technical issue. How do you know you are negotiating a turn across traffic (left turn in the US) with the oncoming Chevy Blazer rather than the similar Chevrolet two blocks down the street?
‘AI is a clearly fad’
Exactly. They programmed a computer to detect coma-shaped formations.
OOOOOOO! ARTIFICIAL INTELLIGENCE !!!
I am also offended by credential inflation: ‘Zheng, a doctoral student.’
She has not earned the right to be called doctor.
‘Now, there is a computer model that can help forecasters recognize potential severe storms more quickly and accurately’
Models are hot, too. This thing isn’t modeling.
These shenanigans tell me to not take this report seriously.
Additionally, I consider the output interesting, but not particularly useful. It reports the formation, or potential formation, of coma-shaped clouds. This phenomena is on a storm front. To whit, meteorologists will ALREADY KNOW ABOUT THE STORM. The computer system telling them it could get severe won’t be news to them. Additionally, what are they going to do with the information? Break into local television to tell people that a tornado could form over New Johnsonville in the next half hour?
First time it doesn’t actually form will be the end of that.
Meteorologists already do a good job of seeing potential troubles in storms. Supercomputers and this program won’t improve that. The system saying, “Hey, watch here!” isn’t going to change anything.
Gamecock. Not that I disagree with you, but I do wonder if this MIGHT not be the first step toward a future two or three decades or so from now where the TV Station Manager fires their last weatherman because the computer does the same stuff just about as well. And it doesn’t need health insurance or ask for raises.
Mammas, don’t let your babies grow up to be weatherpeople?
“…I am also offended by credential inflation: āZheng, a doctoral student.ā
She has not earned the right to be called doctor…”
She was not called “doctor.”
She wasn’t called a graduate student.
You get to be called a doctor when you have graduated, not before.
And my point was that this crap is done to boost the esteem of the “scientist,” to make the story more believable. It’s a cheap trick.
Using NVIDIA or AMD Radeon computational accelerators, today, local broadcasters can have computational systems in a pedestal sized rack that equal the computational power of full building supercomputers of 15 years ago. For local weather alerts they don’t need to run through all of the images of the world, just those locally. With 50 – 100 Teraflops of computational power, doable within a useful time window.
Perhaps AI is the answer too having accurate temperature figures.
At least A,I does not have any ideological baggage to alter the figures.
Keep the humans away from the results, just broadcast them directly.
MJE VK5EL
AI hasnāt worked out so well for the 737 Max in itās early version. Makes nuclear energy look hundreds of times safer than air travel.
Farmer,
there is no any AI pilots certified anywhere in public air travel.
So how can you blame AI for any human or otherwise error in public air travel,
is a bit to far gone…
where no any AI yet involved in piloting,
according to all public knowledge.
Even the case story of “live” AI pilot “training” has to be considered as a very closed covered up, in case of public air travel… not openly publicly declared yet…
if even that, at the very least, where considered as some AI and piloting already there
in action.. and not just a fictional wishful story!
Oh, maybe, that’s what you were looking for! š
Do you know of any AI piloting in public air travel, either simply as in “training’, or maybe as certified piloting??
Please do share, if you happen to know.
cheers
The 737 MAX 800 program was written specifically to detect and respond to certain instrument readings. Certain combinations would indicate a progressing, imminent stall. The program was written to respond by pushing the nose down.
It appears the program was faulty, as was the instrumentation. Many of the planes were sold with only one Angle of Attach vane. If the vane, easily damage by even a relatively small bird strike, indicated a stall the Maneuvering Characteristics Augmentation System, or MCAS would read a stall and try and correct it.
Every plane should have been fitted with dual AOA vanes that could detect a malfunctioning vane.
The MCAS responded too much, too quickly, surprising the pilots. Two problems here. The “too much” response made it vary difficult to over-ride using the controls. The MCAS apparently responded more strongly each time the problem wasn’t corrected.
The pilots never received training in how to shut off the MCAS system, a pair of switches handy to both the pilot and first officer. Some of them found, heard through the grapevine, or discovered for themselves how the switches worked in order to stop the problem and retrim the plane manually.
Boeing down-played the need for additional training in the 800MAX because that would have resulted in a new certification and an expensive retraining class for 800 MAX pilots and crew.
There were a number of other difficulties with how the design, certification and deployment of the plane was done implicating both the FAA and Boing, along with some of the airlines, in how the problem unfolded.
Philo
July 3, 2019 at 6:16 pm
————————-
Philo,
thanks for your very clear and through explanation.
All systems you consider there as in the means of explanation,
are no AI, are simply computerized advanced or even smart ones,
within the realm or environment of the auto piloting…
No AI perse… what so ever
But the way Farmer’s comment above indicates, is that already there maybe AI’s being “installed” in public air travel, without much of any information or transparency there, kinda of obscured from public.
So information about AIs installed in 737 MAX complicates the situation…
regardless of what the real problems and errors with 737 MAX, which definitely nothing to do with AI…
if that attempted information as probably indicated has some “meat in it’s bone”.
So the point here is, as per conversation triggered by Farmer:
“Any AIs there really installed in the 737 MAX!?….
or maybe anywhere else, as far as public air travel concerned”
š
Maybe right time to pose such a question!
Or, as otherwise, also considering the right time to pose a question to the host of this website;
“With all do respect, do you sir, have any clue why your website is deteriorating and so poorly performing in the means of traffic, against any proper odds there???”
((where “proper odds” meaning the extraordinary expected services and support from the Cloud (and maybe its service AIs…))
cheers
Whiten – Iām no AI expert but the MAX program works behind the scenes to make the aircraft handle similar to the earlier 737s so the pilots wouldnāt need a new type rating. Check out Jaun Browneās YouTube channel as heās put together a series of videos on the 737 MAX that are very informative. Itās the blancolirio channel and Juan is a 1st officer on the 777. The AI analogy to the 737 MAX shows that any program, AI or otherwise is only as good as the those programming, reviewing, and certifying the system. Regarding the nuclear analogy in my comment, Iāve worked in DoE nuclear facilities for 13 yrs and can attest to the safety.
Well if it is AI pattern recognition it is not a “computer model”. More media studies student interns trying to write about science.
The US Military (Air Force) has been employing āpattern recognitionā computer programs since the early 1970ās, that Iām aware of.
There was a company located in Rome, NY, named PAR Inc. (Pattern Area Recognition) that produced said software.
As I’m sure you know it IS pattern recognition, trained to recognize patterns which humans can also detect. As such, it simply replaces humans staring at screens. It is not a model and has nothing to do with climate change or global warming, though the media want people to believe otherwise.
ātrained to recognize patternsā
Actually, it should be stated as ā¦ā¦ āprogrammed to recognize patternsā ā¦. by comparing stored āimagesā to newly acquired āimagesā.
During the Falklands War in 1982, the British Navy neglected to āchangeā their on-board āpattern recognition programā and when the Argentinians fired a British made missile at them the software āsaidā ā¦. Donāt shoot it down, its one of ours. A costly decision that was.
“At least A,I does not have any ideological baggage to alter the figures.”
If only.
Unfortunately, one of the (many) problems with AI is turning out to be that the choice of training materials seems sometimes to be inadvertently biasing the results.
I read as far as “Penn state” – I have better things to do with my time than read anything from that institution.
I misread the line that said Penn State AccuWeather inc has developed a program to spot commas, for some reason, I thought it said they can now spot “conman”.
Hey ho, better get down to Specsavers…
They would not have to look very far. Penn State is home to one of the worldās preeminent Climate conman.
1) Greens fill new Weather AI with Climate ‘Data’, like increasing Extreme Weather, Ice Free Arctic, Adjusted Temperatures, ect.
2) Weather AI unable to make useful predictions. All storms predicted to be permanent hurricanes. Simultaneous drought and floods. 90Ā° Blizzards. Ect.
3) Climate Faithful declare It’s Worse Then We Though. Climate Crises has made Weather Unpredictable. Children just won’t know what Forcasting is.
~Āæ~
Hi there my dear colleague and the most indefatigable collaborator
Since our models are not exactly brilliant predicting the climate change, what about this ‘space weather’ lark, do you think there is any money in it?
perhaps good time to jump on an early (gravy) train?
regards (…..)
“perhaps good time to jump on an early (gravy) train?”
Sounds like a valid concept to me. Storyline is — Come the next Carrington Event all our fancy electronics will fail and we’re all gonna die.
Only question is how you are going to monetarize the situation.
Another announcement of success, that appears based upon minimalist success parameters.
No listing of verified storm sightings.
No listing of severe weather warnings that save lives or property.
No listing of data input requirements, or is this all based upon images?
Aggregated data inputs, program coding, super computer operational time and unstated machine outputs finally achieve that lofty achievement; i.e. they might equal or slightly exceed a trained experienced meteorologist’s skills…
Accuweather states they can “detect some comma-shaped clouds before they are fully formed, and our detections are sometimes earlier than human eye recognition”
Maybe…
Dismiss all of those pesky weather forecasters who closely observe Earth’s weather so that they may better forecast weather!
Obviously Accuweather is announcing their “some” and “sometimes” success for honorable humanitarian reasons…
University of AlmerĆa in Spain.ā
No need for forecasting in Spain — the rain there will always fall on the plain. š
Are you sure, just enjoy it:
There was a time when ‘AI’ was commonly understood to be a computer that can think.
Now it means anything but that.
Ah ha! A new metric to say that AGW is getting worse! A rise in comma clouds! Think of all the records that will be set!
Optimizing and heuristic algorithms. Has any AI demonstrated the degrees of freedom exhibited by a human mind?
Hereās what they need to do. Dig up grammas brain. She would look off in the distant clouds and say, āstorms comminā. And sure enough we would have a storm in about 1/2 an hour.
They are obviously using comma delimited data files. š
Supercomputers and computer-learning programs can greatly enhance weather prediction — but only because weather takes place so quickly that it is possible for computers to arrive at solutions before the non-linear equations throw out chaotic results which diverge rapidly from any semblance of reality — and do so in an entirely unpredictable way.
New Zealand is primarily an agricultural country. Milk and milk products are a big part of that. Cows need to be pregnant to produce milk.
When I see “AI” I first think of the main use of the term here, Artificial Insemination!
Possibly more practical than the computer version.
I was in st martin when Irma was approaching . The european model showed a direct hit several days in advance while the US models showed it re curving and going north of st martin almost til the time the eye went over st martin . For whatever reason the US models continue to lag behind the european computers in hurricane forecasting .
Image recognition systems are very good at this kind of thing. They are very thorough and never get tired or bored. I see no reason why this won’t work or why it isn’t a good idea.
This article doesn’t say what kind of lead time this tool provides, except to say “real time.” I would take that to mean 0 to 6 hours before the first report of severe weather, which is about the limit to usefulness for humans to analyze and extrapolate the information from real-time tools like radar and satellite imagery. If a meteorologist on duty doesn’t know that severe weather is possible in that time frame in all or part of his or her area of responsibility, he or she has no business sitting in the hot seat. Comma-shaped clouds may or may not have a portent for severe weather in the succeeding 12 hours. There’s lots of other parameters that factor into severe weather. While it might be nice to have a little confirmation that an area may spawn a cluster of thunderstorms, from my experience I would not think that’s such a valuable addition to the toolkit. But hey, what do I know, I only had 3.5 decades of experience, which doesn’t seem to count for anything at NWS/NOAA anymore. It’s obvious from the changes going on in the operational offices that the management intends to rid the offices of most people and replace them with algorithm-derived output. Doctoral students and even PhD professors who have never issued a forecast or a warning seem to be very trendy in research; some results might be useful, others less so. Allowing meteorologist interns with little to no experience to work radar shifts to issue warnings seems to be trendy too…no wonder some significant events are handled less well than in years’ past. And don’t get me started on the grids. Grrrrrr. Good luck!
“Using artificial intelligence to better predict severe weather.”
Congrats – everyone believes you when you show them!