Polar Wind Data Blow New Life Into Forecasts

Like the rest of their colleagues at weather prediction centers around the world, the weather modelers at the European Centre for Medium-Range Weather Forecasts (ECMWF) have a problem. They’re a little short. Well, more than a little short, says the head of the center’s satellite research section, Jean-Noël Thépaut. In reality, he says, they are probably a few million short.

Thépaut isn’t talking about the center’s budget. He’s talking about the group’s medium-range (3- to 10-day) weather-forecasting model and the number of additional observations of the atmosphere they would need to fill the observational gaps with data.

  Page 2 Photograph of Sunrise and Clouds in the Arctic

Satellite pictures of Arctic clouds are helping forecasters predict the weather in Northern Europe and other high-latitude areas. (Photograph courtesy NASA JPL Microwave Temperature Profiling program)

  Photograph of Jim Key Standing beside an Isolated Polar Weather Station

“Every time we run our global model—every 12 hours—we estimate the current state of the atmosphere using about eight million data points. Typically, we have only two to three million observations to work with over a 12-hour time period, with a large redundancy among some observations” says Thépaut. To make up the missing information, they rely on the values predicted by the previous forecast. In other words, a large part of the model’s input is based on the previous model run’s output. Thus, errors can snowball in the model’s subsequent output.


Weather forecasts in polar regions rely on remote, widely spaced weather stations. Forecasters are looking for new ways to fill in the gaps of missing data, hoping to improve weather predictions. (Photograph courtesy Jeff Key)

  Map of Winds in the Antarctic

In 2003, modeling centers around the world got a little help in filling some of the observational holes in their models when Jeff Key and a few of his colleagues at the Cooperative Institute for Meteorological Satellite Studies (CIMSS) in Madison, Wisconsin, began calculating the speed, height, and direction of the wind in polar regions using imagery collected by NASA’s Terra and Aqua satellites. Thépaut and his colleague Niels Bormann discovered that plugging even a tiny percentage of those observational holes with real-world atmospheric data extends the window of accuracy of their medium-range weather forecasts about three hours—a significant improvement in the world of weather prediction.

  Key for Wind Map

Observations from polar-orbiting satellites augment the coverage of ground-based weather stations around the poles. Scientists use the satellite data to track the motion of clouds and water vapor, which corresponds to wind speed. The map at left shows wind speed and direction at high (yellow), medium (orange), and low (red) altitudes over Antarctica. (Map courtesy NOAA/NESDIS Polar Winds)


New Teams, New Ideas

  page 1Page 3

In summer of 2000, Jeff Key was settling in to his new job leading a polar remote sensing group at CIMSS, an organization that brings together scientists from the University of Madison-Wisconsin, NASA, and the National Oceanic and Atmospheric Administration (NOAA). In July, a senior colleague suggested to Key that one of the sensors flying on NASA’s new Terra satellite was capturing detailed imagery of the polar regions on a regular enough basis to be able to estimate wind speed and direction by how far clouds and water vapor features moved between one overpass of the satellite and the next.

“I was new with NOAA at the time, and my experience was not really with winds,” says Key, “but we did have someone, Chris Velden, who had extensive experience with estimating winds from geostationary satellites. So we thought, ‘Why not give it a try?’”


The group began collecting imagery from a sensor known as MODIS, short for Moderate Resolution Imaging Spectroradiometer. Terra takes about 100 minutes to make the trip from the North Pole, around the Earth, and back, explains Key. “That means we are looking at images captured more than an hour and a half apart.” To calculate winds, they have to be able to see some movement of the clouds between one image and the next, but if too much time elapses, the clouds might not be recognizable any more. “If you go outside and watch clouds in your yard, you can see how fast they can change in minutes. In a one-and-half-hour period clouds can move and change shape a lot. The first thing we had to figure out then was if we could recognize a cloud feature from one image in an image captured 100 minutes later.”


The Moderate Resolution Imaging Spectroradiometer (MODIS) instruments aboard the Terra and Aqua satellites each pass near the North Pole every 100 minutes. This true-color image from August 27, 2003, shows swirls of cloud over the Arctic Ocean. (NASA image by Jesse Allen, based on MODIS data)

  Diagram of Overlap in Coverage of Consecutive Satellite Overpasses

“We did some manual tracking—just pulling images up on the computer and picking out features that we would map changes in. Even with 100 minutes between scenes, we found features that we could follow across several images,” says Key. Based on Velden’s past experience estimating winds from geostationary satellite observations, Key, Velden, and Dave Santek, another CIMSS scientist, developed a similar approach for the Terra MODIS images. Santek modified the existing software that automatically finds and tracks important cloud and water vapor features in geostationary satellite image sequences for use with successive orbits of the polar-orbiting Terra satellite.

After about nine months of development, says Key, they had an automated process running 24 hours a day at CIMSS. Every five minutes, their computer in Wisconsin reaches out over the Internet and connects with a NOAA computer housed at NASA’s Goddard Space Flight Center in Greenbelt, Maryland, to check for new MODIS data.

Once the computer has three images from successive orbits it begins inter-comparing them, identifying cloud and water vapor features and mapping their movement. The program churns out thousands of wind vectors—height, speed, and direction—for each MODIS image. “At first we were just using Terra MODIS data, but in November 2002, we picked up Aqua MODIS data as well. Since the two satellites have staggered orbits, we get imagery from a much more complete area. Plus, using both satellites reduces the time between observations from 100 minutes to just minutes or tens of minutes.”


Polar-orbiting satellites circle the Earth while the planet turns beneath them. The illustration at left shows the ground area observed by MODIS aboard Terra on three consecutive overpasses. Near the Pole, the orbits overlap every 100 minutes or so. Computer analysis of the data in areas where at least three consecutive images coincide reveals wind speed and direction. (NASA illustration by Robert Simmon)

Satellite Image Sequence Showing Cloud Motion Used to Derive Wind Speed

Key and his group are focused on time because the people who need observations of polar wind need it, well, now, explains Key. “We provide the MODIS polar winds to weather modelers all over the world. A number of those centers tell us they need winds that are no more than 3 hours old, which is quite a challenge when you are talking about information that is derived from satellites like Terra or Aqua that were designed for research rather than real-time, operational use.”

Among the first users of the MODIS wind data was the European Centre for Medium-range Weather Forecasts, located in the United Kingdom. It was a time of new beginnings at ECMWF, too. Atmospheric scientist Jean-Noel Thépaut had just taken over the lead role of the group that integrated satellite data sets into the center’s forecast models, and fellow scientist Niels Bormann had just joined the team as well.

“We need these partnerships with modelers,” says Key, “to gauge the quality of the products. The best validation of our work is if the winds have a positive impact on forecasts, but that isn’t something we can determine ourselves. That’s why we need the NWP [Numerical Weather Prediction] centers.”

“There is a long history of collaboration between CIMSS and ECMWF,” adds Thépaut, “This project, though, started informally. I had just arrived as the new head of the satellite group, and Niels was new as well. We went to visit CIMSS, just to introduce ourselves and to get contacts, and they [the MODIS winds group] said, ‘Hey, do you want to try this new polar winds product?’ And we said, ‘Sure, we’ll give it a try.’ But I think it is fair to say we didn’t have very high hopes.”


Thermal infrared (left) and water vapor (right) data reveal structure in the atmosphere and provide estimates of winds. The arrows in these images point to features that can be tracked from one orbit to another (vertical sequence). The changing position of the arrowheads follows the changing location of the cloud. (NASA images by Robert Simmon and Jesse Allen, based on MODIS data)


A Demanding Model

  page 2Page 4

Today’s Numerical Weather Prediction (NWP) models are mind-bogglingly complex. In the models, the basic physical laws by which the atmosphere operates—things like Newton’s first law of motion (force equals mass times acceleration), and the law of conservation of energy (energy cannot be created or destroyed, it can only change form)—are expressed as an interrelated web of mathematical equations. These equations describe the state of the atmosphere—everything from temperature, to humidity, to wind speed, to convection—at a given point in time for tens of millions of locations in the atmosphere.

“To predict the future state of the atmosphere, we have to feed the model with information on the current state of the atmosphere,” says Bormann, which is where observations like the MODIS polar winds come into the story. Determining the current state of the atmosphere is itself a complicated task. Meteorologists today have to find a way to deal with the hundreds of thousands of observations now available to them through satellites, ground-based weather stations, ocean buoys, aircraft, and weather balloons.

  Map of Arctic Weather Stations

During the 12 hours between model forecast runs at ECMWF, other computer programs churn away at the task of bringing together, or assimilating, the information from all these sources into a single batch of information on the atmosphere’s current condition. The results of the assimilation serve as the initial conditions for the next model forecast.

This ‘state-of-the-atmosphere analysis,’ which the modelers call simply the analysis, involves a cycle of comparisons between forecasted and actual atmospheric conditions at various intervals during the 12-hour data assimilation window. “At the start of the 12-hour cycle, we create a short-term forecast based on the atmospheric conditions from the previous analysis,” explains Bormann. “We compare that short-term forecast to the new observations that are coming in, and look for differences. We use the differences that we see to go back and make adjustments to the previous forecast. This is done for all the observations we have for the time window.” At the end of the 12-hour period, all the adjustments are added together to produce a new analysis, a new set of initial conditions to put into the model.

But Why Not Start Fresh?

Why do they use 12 hours worth of observations instead of just using only the most recent observations available at the starting time of the forecast run? Scott Braun, a weather modeler at NASA’s Goddard Space Flight Center, explains the reason with an analogy. “Imagine the state of the atmosphere at a certain point in time as a cork bobbing straight up and down on the surface of the water. If you use only observations from one point in time, then all you know about that cork’s behavior is what you see right then. Imagine if, before you started looking at the cork, someone had reached beneath the surface and pulled it down underwater and then released it. The cork would have bobbed about wildly for a while before settling down into the pattern you observed, but you would never have known it.”

Who knows how that cork might have influenced its surroundings while it was bobbing wildly about? It’s the same with the atmosphere. It makes a difference to the weather prediction model whether the current atmospheric conditions that are about to be fed into the model result from a stable, uneventful set of atmospheric conditions, or whether it is merely a snapshot in time of a highly turbulent atmosphere in a state of flux. In other words, when you want to get the best estimate of the state of the atmosphere, it is important to know how the atmosphere behaved over some time span.


Satellite observations augment the direct observations made by weather stations scattered across the Arctic. The remote sensing data fill in the gaps, both between stations and between model runs. (Map by Robert Simmon, based on data provided by Jeff Key and Online Map Creation)

  Arctic Winds Derived from Aqua MODIS

It makes sense to use all the observations that come in the 12 hours between model runs, after all, those are the “real” data. But why should the modelers bother to go back and adjust the previous forecast? Why don’t they just cut their losses, and start over fresh every 12 hours? They don’t because all the observations available at the time of any forecast aren’t enough to fill the model’s incredible appetite for input. The weather prediction model wants up to eight million data points on the state of the atmosphere at any given time in order to predict how those characteristics will interact with one other to create the next day’s or week’s weather. What do the modelers do when there aren’t enough real-world observations to fill all the slots? They use the predicted values from previous forecasts. In other words, they use the model to feed the model.

“If we were not using the old forecast, we would be throwing away information on the atmospheric state, and that wouldn’t be very smart if you’re already a little short,” jokes Bormann. Instead, it makes more sense to tweak the previous forecast and use the updated predictions to fill in as many observational holes in the model as they can. The value of MODIS data from CIMSS was that they filled a lot of the observational holes in the part of the model that deals with polar winds.

  Key for Wind Map

MODIS-based estimates of wind speed and direction provide input for models where direct observations are lacking. This image shows MODIS-based wind estimates at low (yellow), medium (orange), and high (red) altitudes over the Arctic. The animations show how the winds at the three altitudes change over the course of the day between December 1 and 2, 2004.

  • Animations:
  • small (1.8 MB QuickTime)
  • large (5.7 MB QuickTime)


A Demanding Model

  page 3

“When we started using the MODIS polar wind products, we saw about a 40 percent increase in the number of upper-atmosphere wind observations that we used to generate the polar wind analysis over the area north of 60 degrees North,” says Bormann. Such a large increase in real-world data on polar winds would be certain to improve forecasts, wouldn’t it?

“There are a lot of factors you have to consider,” explains Bormann. “Over the Northern Hemisphere, we have a lot of conventional observations—weather stations, aircraft, et cetera—and traditionally, it is more difficult to get improvements in the models with satellite data because of that.”

In some situations, the accuracy of the forecasts is more limited by the model’s ability to correctly represent atmospheric processes in mathematical terms than it is by lack of real-world observations for the state-of-the-atmosphere analysis. In addition, wind information derived from satellites just isn’t as accurate as wind information gathered from weather balloons or aircraft. In the case of the MODIS winds, the modelers knew it was possible that they would find that it was better to have fewer, but more accurate observations than more, but questionably accurate, observations.

“When the CIMSS group asked us whether we would be interested in their polar satellite winds,” says Bormann, “we were very keen to try these observations in our system. However, there was also some skepticism, as these observations are fairly indirect, and based on many assumptions. They are based on tracking clouds across orbits, assuming that cloud motions represent the wind, and assuming that the appropriate height for this wind can be estimated from the cloud information in the satellite imagery. Such indirect measurements have errors and biases that are difficult to account for in the analysis systems, as we had seen before with wind observations from geostationary satellites. We certainly did not scoff at the idea that these satellite winds could be helpful, but I think we also did not expect quite such a positive impact.”

Re-predicting the Weather

To test the impact of the MODIS polar winds, Thépaut and Bormann did a ‘what-if?’ experiment. They went back into their data “warehouse” of old forecasts and asked the question: would the forecasts made over a particular period have been better on average if they had included MODIS polar winds data?

The two scientists re-created 58 ten-day forecasts, this time using MODIS winds along with all the other atmospheric observations that were used in the original analysis. For each day of the 10-day forecasts, they compared what the model predicted for that day to the analysis of actual atmospheric conditions on that day. They did the same comparison for the original forecasts, made without the MODIS winds.

Over the first five days of the Northern Hemisphere forecast, the average quality of the forecasts made with MODIS winds and those without wasn’t very different. But somewhere around day 4 or 5 of the forecast, that began to change. The old forecast, the one made without the MODIS winds, began to miss the mark faster than the ones made with the MODIS winds: the old forecasts began to deviate from the actual atmospheric conditions sooner. By day 7 or 8, the MODIS winds had added another three hours to the useful “shelf-life” of weather forecasts for Europe and the North Atlantic.

Graph of Data Comparing Forecasts with and without MODIS Data

In the world of weather modeling, a 3-hour extension to the time period over which a forecast is still useful is a pretty big improvement. “In conceptual terms,” says Thépaut, “think of it like this. Typically the quality of a 5-day forecast in the Northern Hemisphere today is about the same quality as the 3-day forecast was in 1981. So, in 22 years time we have improved the forecast by 2 days. [Or about 2 hours a year.] And that is the result of all the people working here at the center, the improvement of the observing and modeling systems, the increase in computing power, et cetera. The improvements from using MODIS winds are about 3 hours or so, which, to a very rough estimate, corresponds to about a year of progress here at the center.”

Of course, he stresses, that is a very approximate kind of estimate, but it goes to show how pleased they are in the improvements the MODIS winds have made. “I think the biggest surprise was just how quickly and easily the whole thing worked and how good the results were—which is probably why other centers have been more eager to try them,” says Thépaut.

The newest member of ECMWF’s satellite products branch, Lueder von Bremen, agrees. “I just got back from the International Winds Workshop in Helsinki, Finland, which is a conference on techniques for getting wind data for weather models, and there were several other NWP centers using the MODIS winds. Nearly all of them were having the same good results as we are.”

Five numerical weather prediction centers are already using the MODIS winds in their operational forecasting system: ECMWF, NASA’s Global Modeling and Assimilation Office, Japan’s Meteorological Agency, the Canadian Meteorological Centre, and the U.S. Navy’s Fleet Numerical Meteorology and Oceanography Center. A sixth center, the United Kingdom Meteorological Office, is using the winds experimentally.

Key and his group at CIMSS will keep refining the MODIS winds products, increasing their accuracy and reducing the delivery time. They are going to make a single MODIS winds product that combines all available Terra and Aqua MODIS observations, to cover the widest possible area in the shortest possible time. They are also looking into getting data from a handful of MODIS “Direct Broadcast” stations—facilities that have their own antenna to receive the MODIS digital data stream—rather than just relying on the NOAA-NASA data feed.

Thépaut will welcome any improvements. “Especially if they could get the delivery time down to an hour,” he says. But he laughs because he knows that’s not very likely given the time between orbits and all the other steps that have to happen. Thépaut is a little uneasy about the future of products like the MODIS winds, however. The upcoming NOAA-NASA-Department of Defense mission called the National Polar-orbiting Operational Environmental Satellite System, or NPOESS for short, is intended to continue collecting many of the types of Earth observation data sets currently being collected by Terra and Aqua.


Scientists compared European forecasts made with (blue line) and without (green line) MODIS wind data to the atmospheric conditions that actually occurred. Disagreement between a forecast and the actual state of the atmosphere is expressed as ‘anomaly correlation.’ The higher the percent, the better the agreement. Through day 5, both types of forecasts were in good agreement with the atmospheric conditions that actually occurred, but by day 7, the forecasts that used the MODIS winds were in better agreement with the actual weather than the forecasts that did not. (Graph courtesy Bormann and Thépaut, ECMWF)

  Comparison of Winds Derived from Thermal Infrared and Water Vapor Data

Unfortunately, says Thépaut, “the planned NPOESS visible/infrared imagers are lacking the spectral channels to detect water vapor features, at least the first two instruments in the series anyway.” Key and his group use water vapor observations along with clouds to detect which way the wind is blowing, and NPOESS will not have that capability. “This is a big concern in the meteorological community. These bands weren’t part of the instrument design because at the time, no one was thinking we would be able to do this with MODIS. We discovered it too late. We know that some efforts are being undertaken in the United States to change the situation for the third satellite in the series, and we can just be hopeful that this will happen. Otherwise, what will we do when there is no more MODIS? We certainly don’t want to lose what we have gained.”


Key and his team use thermal infrared and water vapor observations from MODIS to estimate polar winds. Some MODIS-like sensors planned for the near future may only collect thermal infrared data, not water vapor. The pair of images shows the difference in the number of wind estimates that can be made at low (yellow), medium (orange), and high (red) altitudes from thermal infrared (top) and water vapor (bottom) data from MODIS.