Predicting Seasonal to Interannual Climate Variations

James L. Kinter III, Ben P. Kirtman, and Paul A. Dirmeyer

Successful predictions of the El Niño event of 1997-1998 up to two seasons in advance inspired great optimism in the research community that studies the predictability of seasonal climate variations. That event, one of the two largest such events in the past century, was marked by sea surface temperatures (SSTs) several degrees warmer than normal over millions of square kilometers in the eastern tropical Pacific that have been linked to large climate disruptions around the globe directly attributable to the El Niño event. These SST anomalies occur primarily through a tropical atmospheric variation called the Southern Oscillation that typically accompanies El Niño and remote teleconnections in the global atmosphere. The effects of El Niño and the Southern Oscillation (ENSO) are felt, for example, in Indonesia, where drought leads to unhealthy quantities of dust in the air; in northern Australia and northeastern Brazil, where the resultant droughts can significantly alter agricultural yields; and in North America, where excessive rains in southern California and across the northern Gulf of Mexico cause flooding and damage to crops and property.
For very large SST anomalies in the tropical Pacific, the resulting pattern of atmospheric circulation is quite predictable, and the associated disruption of rainfall is found even far away from the tropical Pacific. A global atmospheric model can well simulate the response to large warm and cold ENSO events (Figure 1).
Day-to-day weather cannot be predicted beyond about 10 days; even the short-term weather forecast is given in probabilities (for example, a 30 percent chance of showers). Nevertheless, the slowly evolving changes in the earth’s surface conditions (SST and land surface properties) enable predictions of the statistical behavior of the climate system over several months, seasons, and perhaps years. By predicting the slowly evolving surface conditions, forecasts can be made about changes in the statistics of day-to-day weather fluctuations. Climate forecasts must therefore be made in a probabilistic framework, and a multitude of model predictions must be combined to provide a good estimate of the probability distribution.
Successfully predicting the largest climate anomaly of the past century, however, does not necessarily translate into a routine prediction capability. Despite a clear need for climate predictions that could be used to prepare for agricultural, economic, and human health impacts, the ability of current prediction methodologies to produce accurate forecasts is quite modest. A combination of factors contribute to this lack of ability, including an incomplete understanding of the relevant physical processes governing seasonal climate variability, the inadequate model resolution used given today’s computer limitations, and the noise contamination of coupled ocean-atmosphere signals in the prediction models.
A two-pronged approach is being used to improve our understanding of the physical climate system and thereby to predict its variations more accurately. On the one hand, better representations of processes, such as cloud formation, thunderstorms, and the interaction between the surface and the atmosphere, are continually being developed in computer models. Also, as computers become faster and as our ability to use hundreds and even thousands of computer processors efficiently improves, we can increase model resolution. As climate models gain better representations of the relevant processes and much finer grid spacing, we can expect to see much greater fidelity in the representation of the earth’s climate.
On the other hand, while these model improvements are under development, we can take advantage of the fact that different models of the earth’s atmosphere and oceans behave differently and have different errors. When forecasts made by multiple models are statistically combined to estimate the probability distribution, it is also possible to filter the errors and to simulate climate more accurately. For example, we know that current computer models of the global atmosphere produce too much climate noise or unpredictable variations that are not related to the underlying surface variability. To address this particular problem, a new methodology has been developed at the Center for Ocean-Land-Atmosphere (COLA) Studies that is based on a different way of coupling the ocean and the atmosphere. The basis of this new methodology is called an interactive ensemble. The interactive ensemble reduces the model-generated atmospheric noise at the air-sea interface that is unrelated to the predicted surface boundary conditions. This methodology significantly improves climate models’ simulation of El Niño and its remote effects.

Land Surface and Climate
Unlike the oceans’ relatively unchanging surface, the land surface is extremely varied. Land surface contains tremendous elevation changes, vegetation ranging from dense forests to bare deserts, and regions of snow, ice caps, glaciers, and wetlands. This diversity makes the land surface difficult to simulate and makes its impacts on climate much more regional in scale. In addition, because humans are constantly altering the land surface through agricultural practices (farming, ranching, irrigation), deforestation, and development, climate fluctuations contain a human element on top of any natural variability. Climate models must take these factors into account.
The physical processes at the land surface are modeled on a wide range of time and spatial scales. Ultimately, the highest possible resolution in land surface imagery and grid spacing in land surface models might be needed to represent these processes accurately. Land surface schemes are a fundamental component of the climate models used to understand the effects of land use change; to understand the feedbacks between climate and surface properties, such as soil moisture, snow cover, and vegetation; and ultimately to predict climate. In particular, changes in the balance of water and heat at the surface of the earth are the keys to understanding the feedbacks between land and atmosphere (Figure 2).
Feedbacks are processes by which the land and the atmosphere mutually affect one another. For example, a prolonged drought might cause vegetation to wither and die, changing the color of the land surface as viewed from above as well as its overall roughness and its ability to evaporate water drawn through the roots of the plants. All of these changes reduce the flow of energy and moisture from the land to the atmosphere, making future storms less intense. Studies of these feedbacks, in the context of their effects on climate, are performed with land surface models with an eye toward improving our ability to predict climate from months to seasons in advance.

Distributed Data Management to Aid Forecast Value Assessment
The new strategies described above — ensembles of model predictions, multiple model prediction systems, interactive ensembles, and using very-high-resolution remote sensing data — all lead to the rapid growth in size and complexity of the datasets that are the digital representations of climate predictions. To make these data useful for applications of climate predictions ranging from water resources management to energy forecasting to agricultural competitiveness, we need to process and to combine data that are generated and stored in distinct physical locations.
For example, suppose that water resources managers and disaster preparedness planners need to estimate the likelihood of a summer 2003 recurrence of the 1993 upper Mississippi River floods. One way of producing such an estimate would be to use several climate models, making an ensemble of forecasts for summer 2003 to produce a probabilistic estimate of the likelihood of the event in question. The predicted probability so obtained could be compared quantitatively with a probabilistic measure of the skill of precipitation simulations in the upper Mississippi catchment for each of the models based on a large sample of historical simulations. The historical simulations could also be used to estimate the uncertainty of the prediction. A further quantitative measure of the value of such a prediction could be computed by means of a decision support model that can ingest probabilistic information and can help decision makers quantify the value of various scenarios.
In this example, the observations of precipitation and stream flow at the highest possible spatial and temporal resolution over a period of at least 20 years, as well as a large ensemble of model simulations of the same quantities, would have to be combined. It is entirely likely that the observational datasets and the model output datasets would reside in different data centers, managed by different agencies or even different countries. To enable the type of analysis described above in a timely manner to make it useful and valuable to the water resources and disaster managers, a distributed data management system that could seamlessly support the analysis of these large, complex, and disparate data sets would be required. A prototype of such a system is being operated in several locations, such as at the COLA Grid Analysis and Display System — Distributed Ocean Data System (GrADS-DODS) server (http://cola8.iges. org:9191/index.html). The capability to rapidly subset and process data being served at several remote locations is critical to this capability (Figure 3).

Summary
Predicting climate on seasonal and longer time scales depends on bringing together many capabilities. The effects of coupled ocean-atmosphere and coupled land-atmosphere processes must be represented in complex, dynamical models of the climate that accept as input the high-resolution, global observations available from satellite-based observing systems. These models must be combined statistically to achieve a probabilistic estimate of predictive skill and forecasts of value. Realizing this potential will require advances in our understanding of the climate system, advances in computing, and advances in distributed data management, all of which are within reach during the next five years.

Acknowledgments
Funding for this research has been provided through the NASA Earth Science Enterprise research program under NASA grant NAG5-11656 and through the NASA Earth Science Information Partnerships under a subcontract from George Mason University.
The assistance of D. Paolino of COLA in producing one of the figures is gratefully acknowledged.

About the Authors
James L. Kinter III is the Executive Director of the Center for Ocean-Land-Atmosphere (COLA) Studies in Calverton, Maryland. He manages the scientific program and administrative infrastructure of COLA and serves as Director of the NOAA Applied Research Center at COLA. Dr. Kinter can be reached at [email protected].
Ben P. Kirtman is an Associate Research Scientist at COLA and an Associate Professor of Climate Dynamics at George Mason University. His primary research interests include climate modeling, prediction and predictability. Dr. Kirtman can be reached at kirtman@ cola.iges.org.
Paul A. Dirmeyer is an Associate Research Scientist at COLA, specializing in land surface modeling and the study of land-climate interactions. Dr. Dirmeyer can be reached at dirmeyer@cola. iges.org.
COLA is a center of excellence in the Institute of Global Environment and Society. COLA’s mission is to explore, establish, and quantify the variability and predictability of Earth’s climate variations on seasonal to decadal time scales through the use of state-of-the-art dynamical coupled ocean, land, atmosphere models and to harvest this predictability for societally beneficial predictions.

Back