COLUMBIA, Mo. – The world’s oceans cover more than 72 percent of the earth’s surface, impact a major part of the carbon cycle, and contribute to variability in global climate and weather patterns. However, accurately predicting the condition of the ocean is limited by current methods. Now, researchers at the University of Missouri have applied complex statistical models to increase the accuracy of ocean forecasting that can influence the ways in which forecasters predict long-range events such as El Nińo and the lower levels of the ocean food chain—one of the world’s largest ecosystems.
“The ocean really is the most important part of the world’s environmental system because of its potential to store carbon and heat, but also because of its ability to influence major atmospheric weather events such as droughts, hurricanes and tornados,” said Chris Wikle, professor of statistics in the MU College of Arts and Science. “At the same time, it is essential in producing a food chain that is a critical part of the world’s fisheries.”
The vastness of the world’s oceans makes predicting its changes a daunting task for oceanographers and climate scientists. Scientists must use direct observations from a limited network of ocean buoys and ships combined with satellite images of various qualities to create physical and biological models of the ocean. Wikle and Ralph Milliff, a senior research associate at the University of Colorado, adopted a statistical “Bayesian hierarchical model” that allows them to combine various sources of information as well as previous scientific knowledge. Their method helped improve the prediction of sea surface temperature extremes and wind fields over the ocean, which impact important features such as the frequency of tornadoes in tornado alley and the distribution of plankton in coastal regions—a critical first stage of the ocean food chain.
“Nate Silver of The New York Times combined various sources of information to understand and better predict the uncertainty associated with elections,” Wikle said. “So much like that, we developed more sophisticated statistical methods to combine various sources of data—satellite images, data from ocean buoys and ships, and scientific experience—to better understand the atmosphere over the ocean and the ocean itself. This led to models that help to better predict the state of the Mediterranean Sea, and the long-lead time prediction of El Nińo and La Nińa. Missouri, like most of the world, is affected by El Nińo and La Nińa (through droughts, floods and tornadoes) and the lowest levels of the food chain affect us all through its effect on Marine fisheries.”
El Nińo is a band of warm ocean water temperatures that periodically develops off the western coast of South America and can cause climatic changes across the Pacific Ocean and the U.S. La Nińa is the counterpart that also affects atmospheric changes throughout the country. Wikle and his fellow researchers feel that, through better statistical methods and models currently in development, a greater understanding of these phenomena and their associated impacts will help forecasters better predict potentially catastrophic events, which will likely be increasingly important as our climate changes.
Wikle’s study, “Uncertainty management in coupled physical-biological lower trophic level ocean ecosystem models,” was funded in part by the National Science Foundation and was published in Oceanography and Statistical Science.
Terry Oldberg saysMarch 23, 2014 at 9:34 am
… Climatologists err in appropriating the ideas of “signal” and “noise” from electrical engineering …
Perhaps a better fit: hypocycloids (circles within circles) with varying parameters constrained as to cycle length, distance, etc. and some parameters ‘offset’ as with lobes on cam.
Roger D PGeol says:
March 23, 2014 at 9:47 am
Exactly. You speak from hard earned experience. Too bad no climate modeler has similar experience. In addition, the geology of reservoirs is a rather rich context compared to the average temperature of the planet.
Oh their aching sacroiliacs! What wonderful phrasing for calling their derrieres dense and pulling numbers from it!
I’ll have to remember that one! And your overall post is excellent even without subtle wordplay.
Tornado watches, indeed many weather related ‘watches’ are issued when the conditions are ‘ripe’ for the weather related phenomena to form.
Minutes can pass between clear skies to towering cumulonimbus clouds starting to form anvils and rumbling ominously. Without interference super cells can form and begin rotating.
The idea of the ‘Tornado watch’ is that one keeps tuned to both the skies and NOAA weather alerts. If one is living in Tornado Alley then there are often local TV/radio stations that will broadcast frequent Doppler alerts along with likely paths.
When a tornado or a tornado cloud is spotted, then ‘tornado watches’ get upgraded to ‘tornado warnings’. If I am in the path of a tornado warning path, we clear (turn off lights, stove, oven…) the house and head for the basement with the weather radio and pets. And I’m not even in Tornado Alley!
But I have been in Tornado Alley and gotten caught in serious conditions that were not there 30 minutes prior, nor were they on the horizon. Just ill luck that I happened to be under where the weather conditions went from ‘ripe’ to unstable to storms. From sunscreen covered to rain jacketed with rain covered glasses and frantically trying to get somewhere lightning was not striking. If I heard a roar, I would’ve been satisfied with a ditch and grass.
Like watching the ‘stormchasers’ who finally caught a ‘tornado’ or got caught by one.
Terry Oldburg: In the construction of a model, I favor a method that is more aptly described as “optimization” than as “estimation.” The model builder optimizes the missing information in each of the inferences that are made by the model by minimizing it or maximizing it under constraints expressing the available. I favor this method because it is entirely logical and produces superior results..
That sounds like the E-M algorithm.
I personally only respect Bayesian methods where there is a demonstrably accurate distribution of measurements in some describable population supporting the prior. That is basically Samaniego’s message: in order to achieve an improvement over MLEs (for example) the prior has to be at least sufficiently accurate: past or above a “threshold” as he puts it.
Matthew Marler:
I’m thinking of entropy minimax pattern discovery as described by Ron Christensen in the book “Multivariate Statistical Modeling.” Christensen uses Bayes’s theorem in solving the inverse problem that the model requires probability values but observational science gives the model builder only frequency values. For this application of Bayes’s theorem, the uninformative prior can be shown to uniquely be the uniform prior on the interval between 0 and 1. That the prior is unique avoids the usually just criticism that is lodged against Bayesian methods by the frequentists while providing a solution to the inverse problem.
Early in the development of entropy minimax pattern discovery, Christensen tried MLE. This resulted in the straight rule. He found, however, that in this context MLE led to catastropic failure of his algorithm. This resulted from a usually slight over estimate of the information content in the observed frequencies by the straight rule.
How can there be better predictions if the science is settled?
” Terry Oldberg says: March 23, 2014 at 9:55 am
Despite the difficulties that you describe, statistically validated models have already been developed that predict average air temperatures over the western states of the United States over a forecasting horizon of 6 months. This has been made possible by the existence of 328 statistically independent observed events extending back to the year 1850.”
I’m sorry but since the climate going forward is meant to be changing, what possible value can historical events be in predicting the future by statistical methods?
son of mulder:
Good questions! Perhaps you’d find an answer by reading the articles at http://judithcurry.com/2010/11/22/principles-of-reasoning-part-i-abstraction/ and http://judithcurry.com/2010/11/25/the-principles-of-reasoning-part-ii-solving-the-problem-of-induction/ .
http://wattsupwiththat.com/2014/03/22/new-statistical-models-could-lead-to-better-predictions-of-ocean-patterns-and-the-impacts-on-weather-climate-and-ecosystems/#comment-1596571
@Mac the Knife
Check the WUWT blog roll for Jennifer Marohasy
http://jennifermarohasy.com/2014/03/mega-drought-for-murray-darling-predicted-by-kevin-long/#comments
Funnily enough, I came here from there.
http://jennifermarohasy.com/2014/03/mega-drought-for-murray-darling-predicted-by-kevin-long/?cp=1#comment-554225
corrected first link
http://jennifermarohasy.com/2014/03/mega-drought-for-murray-darling-predicted-by-kevin-long/
“Bayesian hierarchical model”
Ah, another computer game, how nice.
Is it out for Xbox One yet?
How does it compare with Game of Thrones?
Terry Oldberg, I’ve read the references that you have quoted to me. I can’t see how the process described differs from predicting, a) tomorrow will be like today, b) summer will be warmer than winter etc. a) has some use if you haven’t heard a forecast for tomorrow. b) is of no added value as we know that.
But what about predicting how the jetstream will behave over Europe in 2044-2054,so we know whether and where to construct bigger and better flood relief and prevention systems or more reservoirs in 30 years time?
son of mulder:
Your examples a) and b) give me a clue to what it is that you don’t understand about the lesson that I’m trying to convey to you. Thanks for giving me the opportunity to clarify.
In the referenced articles I derive logical principles which, when followed, may result in the ability of a decision maker to predict the outcomes of events lying in the future though this decision maker has only observed the outcomes of events lying in the past. To do so is to “generalize from specific instances.” How to do so is “the problem of induction.” Solving this problem is the topic of the second article. Generalization is facilitated by abstraction; this is the topic of the first article.
Following the logical principles results in construction of a model, one of whose functions is to make a predictive inference. A “predictive inference” is an extrapolation from an observed state of nature to an unobserved but observable state of nature. The former state is called the “condition” while the latter state is called the “outcome.”
Under circumstances found in practice, predictive inferences consistent with observed events are of infinite number but a decision maker must select one of them in making a prediction. How can he make this selection? This is where the logical principles come in. They select that predictive inference which is most useful for decision making from among the many possibilities. The most useful predictive inference is the one that provides the decision maker with the most information about the outcomes of the events of the future.
In practice, the information provided by the model is incomplete with the result that the model predicts the outcome of each event only to within a probability. Neither of your examples illustrates the idea of imperfect information or probability.
A predictive inference references the set of all possible conditions and the set of all possible outcomes. Neither of your two examples references the two sets. Also, in selecting that predictive inference which provides the most information, one varies the definitions of the conditions. In composing your examples, you’ve not done that.
Regarding the usefulness of the predictions, this property is imparted to the model through the selection of the set of all possible outcomes. This set should be selected by reference to the decisions that will be made by the model. In the case of a trial of a lung cancer cancer drug, it might be appropriate to select the set { alive, dead } in reference to a patient a year after he has taken the new drug. For a lung cancer patient to know the value of the probability that he’ll be dead in a year conditional on whether he takes the new drug will be useful to him.
In coming up to speed on the topic of my tutorial, you’d find it useful to compose an example featuring a set of conditions and set of outcomes. wherein the definitions of the outcomes are constants but the definitions of the conditions are variables. You should select the definitions of the outcomes for their pertinence to making a decision that is important to someone. The conditions and conditional outcomes should have probabilities of occurrence. A model that is described in this way is amenable to optimization through use of ideas from information theory.
tSteven Mosher says:
March 22, 2014 at 1:41 pm
Typically enough, you fail to recognize the difference between what needs to be done scientifically to unravel a physical problem–which is what I address–and what someone claims to be an improvement in purely phenomenological statistics. I don’t have time for such uncomprehending quarrel.
Terry, you are asking for a lot of time to be invested in a detailed study of what I glean to be a fusion of Bayesian Probability Theory, reformulation of systems into phase spaces and a sort of (analogous to) a Least Action principle to choose the best prediction. But before that you need to show me how your methods would practically address the 30 year European jetstream prediction because that is essentially what determines storm patterns in Europe, in a context where there is a continuing change in the environment (anthropogenic CO2) and a naturally chaotically evolving physical system. That would give me some idea how practical it is to apply what you are proposing on a fairly basic climate scenario to give a useful answer.
I’m afraid my mathematical instinct is that because the world climate is an open chaotic Thermodynamic system the same lack of convergence that one sees in say the closed classical 3 body problem will be present in predictions of the jetstream’s behaviour but more so.
son of mulder:
The topic that you’d need to study, if you have not already done so, is logic. As currently organized, global warming climatology is illogical.