Because we cannot directly observe these things, forest ecologist Steven Running and his research team at the University of Montana constructed a computer model of the Earth's terrestrial biosphere that he expects will teach us a great deal about the dynamic interactions between land plants and the lower atmosphere. The model enables him to watch as changes in the biosphere literally ripple across continents in response to meteorological changes (Ford et al. 1994).

"We can watch daily photosynthetic activity as plants react to changes in weather patterns and storm tracks going across continents," Running says. "Suppose it is in the middle of the growing season. A storm comes through that drops temperatures from the 70°F to the 50°F range, and the skies go from sunny to cloudy. The biosphere will react by dropping its rate of photosynthesis. We can see this response sweep across the landscape over a period of hours."

What is a computer model?
Elissa Levine, a soil scientist at NASA's Goddard Space Flight Center, explains that a computer model is a numerical representation created to simulate a real system. "We use models," Levine says, "because we don't always have access to the real system at the desired scales of time and space."

According to Levine, scientists don't have enough information about the Earth's processes to understand how they function as a whole, integrated system over time. We know that ecosystems vary greatly from local to global scales. Yet, in order to characterize how these ecosystems function, scientists need to generalize across each region (Weishampel et al. 1999). Scientists put the information they know about a system--as either numerical measurements, or theoretical estimates--into a model and in turn the model tells them two things: (1) how the inputs respond to each other, and (2) where there are gaps in their knowledge that need to be filled.

Levine works with two types of models: (1) simulation models and (2) data-driven models. Simulation models use equations that are written into a program that acts as a copy of a real system. Inputs that make the model run are things like site conditions (soil properties, land cover, slope, aspect, latitude, longitude, etc.), temperature, and precipitation values. The types of simulation models she works with are at the "process" level, which means they look at the system at the local level that gives details of how specific mechanisms work. The results of these models can then be generalized to larger, coarser scales (Knox et al. 1997). When the model is run, the equations work together to make a prediction about a certain part of the ecosystem that is being simulated (such as the amount of moisture in the soil at a certain depth and time, the amount of photosynthesis occurring at a certain time, etc.).
 

These three maps of North America show the amount of carbon (in the form of carbon dioxide) used by plants over a one-year period. This is called net primary productivity, and is related to the growth rates of green vegetation. Understanding the exchange of carbon dioxide between the biosphere and the atmosphere is critical to predicting global climate change.

Each image is based on the output of a different computer model, each of which used current climate conditions and carbon dioxide concentrations to calculate plant productivity. (Image courtesy of the University of Montana Numerical Terradynamic Simulation Group)

Computer models can fill in the blanks left by satellite remote sensing and in-situ measurements. Scientists at the University of Montana produced this animation of daily net primary productivity (carbon uptake by plants) from model results. Green represents carbon being consumed by vegetation (positive productivity), while tan shades are carbon being released back into the atmosphere (negative productivity). (Image by Robert Simmon, based on data provided by the University of Montana Numerical Terradynamic Simulation Group)

The second type of model uses observed measurements that scientists have collected across the globe to generalize real processes and properties seen in nature. Levine uses Geographic Information System (GIS) software tools to combine maps and other types of data from satellite sensors (such as land cover, soil types, and climate data). Then, she applies models that use techniques such as statistics or "neural networks" to interpret the real relationships between each of the different layers of data. [A neural network is a system of multiple computer processors connected together so that they share information via a common program that enables them to "learn" by trial and error in a manner similar to the way the human brain works.]

For example, Levine uses satellite images of land surface overlain on soil data and, together with the GIS software and a neural network, relates soil properties with the spectral signatures from the satellite image. These signatures are colors of light that are reflected (both visible and infrared light) by the vegetation and other objects on the ground. Using the relationships that are derived from this information, Levine’s model can make predictions about soil carbon and other properties for areas where there are little or no data (Levine and Kimes 1997).

When the model predicts something interesting or unexpected, Levine says, then it is time to go make more measurements.

Why Build Computer Models?

The data used in this study are available in one or more of NASA's Earth Science Data Centers.

Data-driven models combine many types of observations to derive information that is difficult or impossible to directly measure. In this case, Dr. Levine used topography, vegetation index, and in-situ soil sample data from selected sites to estimate soil carbon content throughout Brazil's Amazon Basin. (Image by Robert Simmon and Steve Fifer, NASA Goddard Space Flight Center.)

Why Build Computer Models?

Computer models are essential for studying complex systems, where many variables are operating simultaneously. Most biology experiments are done by holding all other factors constant while studying one variable. But the controls, constraints, and behavior of complex systems cannot be studied that way (Waring and Running 1998). Running recounts a previous experience in which one of his models predicted a system-level behavior that he later confirmed through direct measurements.

"In 1974, we had never measured the fluxes of water in really big trees like redwoods or Douglas firs," he recalls. "We created a computer model and entered all we knew then about the transpiration rates and water balance of those trees." (Water balance refers to the amount of water stored within a tree after adding the amount of water taken up through the roots and subtracting the amount evaporated from the leaves during transpiration.)

"The models showed us that those trees could not have the same transpiration rates per unit leaf area as small trees, or they would die by mid-summer. The model suggested that large trees had transpiration rates almost 10 times lower than small trees. This had never been measured before. When we went to the field years later and made the measurements, we found that the model was correct. This is the first example in my career where a model taught me something about the ecosystem that I didn’t previously know." (Running et al. 1975)

Running’s current global model allows his team to visualize changes in the biosphere at resolutions no real data set can provide. At time scales ranging from minutes to months, and at spatial scales ranging from acres to the entire Earth, Running and his colleagues can toggle certain variables and then run a scenario over and over, watching and learning as their virtual creation plays out possibilities.

In a matter of seconds to minutes, Running’s team uses the model to simulate interactions that in reality would take them years to measure. For instance, they can use the model to examine what causes changes in the beginning date and length of growing seasons (Waring and Running 1998). We know that growing seasons vary from one year to the next. Sometimes the differences are subtle, sometimes extreme. There are climate anomalies such as El Niño and La Niña that can also dramatically affect a region's growing season. Witness some of the record high temperatures recorded in North America during the 1997-98 El Niño, followed by the most severe drought ever in the mid-Atlantic United States during the 1998-99 La Niña.

Using current satellite sensors, Running's team measures many of the same variables that they model, but the satellite data are integrated into 14-day "composites" (meaning that one image represents the average values measured over a 14-day period). But suppose something interesting happens in the middle of those 14 days? Without models, Running's team could miss important details.
 

This is a simple diagram of the model used by Steve Running to predict the rate at which large douglas fir trees exchange water with the air. The model incorporated exchange of water from the soil to the roots, storage of water in the root system and tree trunk, and release of water from the needles to the atmosphere.

Additionally, Running points out that models allow his team to compute variables that cannot be directly measured. "We can only directly measure photosynthesis for an individual plant," he explains. "But models allow us to compute that variable on a global scale."

Models provide the only means for comprehensively examining the terrestrial biosphere. It is difficult to measure one variable at one scale of time and space (such as the photosynthetic activity of a single plant), and then integrate that measurement with other measurements at other scales of time and space (such as changes in temperature and precipitation). Models allow scientists to integrate multiple measurements across varying scales of time and space into a single tool for visualizing the system and predicting future changes.

"A mature, well-tested model can provide a prediction capability that data alone cannot provide," Running states. "There is no way you can look at data and extrapolate into the future. The only way we can imagine what the terrestrial biosphere may be doing 10 or 100 years from now is through modeling."

How to build better climate models
Modeling Earth's Land Biosphere

Modeled soil temperature (animation) for a single day of 1984. Blues are cold temperatures (below freezing) and reds are warm, while white indicates zero degrees celsius. Because plants stop photosynthesizing when the ground is frozen, soil temperature is an important factor in determining net primary productivity. (Image by Robert Simmon, based on data provided by the University of Montana Numerical Terradynamic Simulation Group)

How to build better climate models

Many Earth system scientists routinely use computer models in their research. Yet they have generally moved beyond the creation of new models, and are working to improve the existing ones. A model isn't finished until it realistically portrays the system it is designed to simulate and accurately predicts how that system will change in the future. The point is that, today, Earth system science is less likely to result in totally new discoveries than it is to gradually revise our understanding of our complex home planet through a deliberate, iterative process (sort of like peeling an onion, one layer at a time).
 

According to Levine, the sophistication of the model reflects the maturity of the science. In the 1980s, when Levine was still a Ph.D. student at Penn State University, she studied the effects of acid rain on soil using a simulation model. Levine used existing soil profile information that described the physical and chemical properties of the soil at various depths. She also used climate information that described the amount and acidity of the precipitation falling on the soil. Then she ran her simulation model to predict the changes in soil chemistry over time, and to identify soils that would be sensitive to acid rain inputs.

Levine found that certain soils are more sensitive to, while others are more tolerant of, acid rain. She describes the sensitive ones as shallow, letting water pass through quickly. Shallow soils typically don’t have much capacity for holding nutrients and they don't contain much organic matter. Conversely, soils that are more tolerant of acid rain are generally thicker, more fertile, contain more organic matter, and filter water through more slowly. In the shallow, sandy, and low buffering soils, Levine found that, over time, acid rain decreases fertility, increases the acidity of the soil, and affects the biological contents (bacteria, worms, etc.) of the soil.

When she came to NASA in 1987, Levine was asked to understand the soils of the entire world. Her challenge: how to extrapolate from digging holes at single sites to mapping the soils of the Earth? The most burning scientific questions did not allow her to look at soils in isolation, but instead to recognize the role soils play in the whole ecosystem. She began working with scientists in other disciplines to identify links between each of the subsystems of the natural environment. (Here, "subsystems" refers to all of the small components within the natural environment that are contained within, and comprise, the terrestrial ecosystem.)
 

This soil profile—an oxisol—is similar to the soils found throughout the Amazon basin. Because directly measuring the organic carbon content (by digging a pit) is a laborious process, scientists use models to extrapolate data. (Photograph courtesy National Soil Survey Center)

"It turns out, the interfaces between Earth’s subsystems are where most of the significant processes are happening," Levine observes. "To understand things like the exchange of carbon dioxide between the soil and the atmosphere, or the uptake of nutrients from the soil by plants, or the return of nutrients through rain and litter back to the soil--to understand all of these things requires us to examine the interfaces between Earth’s subsystems."
 

The ground-based soil sampling sites shown here (red dots), scattered throughout Brazil's Amazon Basin, can not provide enough data to map the entire region. However, these data are used to "train" computer models to recognize patterns in satellite imagery that relate to conditions on the surface that can not otherwise be measured remotely. (Image by Steve Fifer, NASA Goddard Space Flight Center)

Today, Levine is using her data-driven modeling approach to map soil and vegetation types across Amazonia, South America. Her goal is to find out how much carbon is stored in the soils there. Using data from 200 sites on the ground in the Brazilian Amazon, Levine and her colleagues "trained" a neural network using soil, land cover data, and satellite imagery. (Levine’s model uses a program that can improve itself, or learn, by comparing its predictions to real measurements and then revising its processing algorithm. The process of preparing a given model to make certain types of predictions is called "training the model," which in this case is Levine’s neural network.) Levine then used the trained model to estimate the amount of soil carbon present in the Amazon region based on satellite imagery alone (Levine and Kimes 1998). Next she is planning to generate soil-carbon maps for a larger portion of the Amazon region--a task that would take many years to complete if she relied solely on ground-based measurements.

The Leap from ‘Local’ to ‘Global’
According to Running, making the transition from small-scale to global-scale models wasn’t easy for many terrestrial ecologists, who were trained to study individual organisms. In the early 1980s, most vegetation models were constructed for individual trees to show things like water and carbon movement. But Running notes that the small-scale models helped scientists identify the ecosystem processes that larger-scale ecosystem models needed to incorporate--things like the carbon cycle, the processes of respiration and photosynthesis, decomposition, canopy characteristics, soil type, regional hydrological and nitrogen cycles, etc.

"In adopting a global perspective, we learned that we need to stop worrying about each plant as an individual," Running explains. "In producing remote sensing data products like the Normalized Difference Vegetation Index (NDVI), we began to think more abstractly about the biosphere as a ‘chlorophyll sponge.’ Land surface cover became an aggregation of all vegetation, not individual plants."
 

This map, derived from a model that combines data from ground stations (top of page) with satellite imagery, shows soil organic carbon content where it is difficult to obtain these measurements directly. (Image by Steve Fifer, NASA Goddard Space Flight Center)

Running is quick to point out that global-scale modelers did not abandon the strategy of direct measurements that got them here. Rather, most modelers use a nested approach in which model results are compared to actual measurements (or validation data) obtained at multiple sites around the globe. Running’s team coordinated the establishment of a worldwide network of towers equipped to measure a range of variables in the atmosphere (temperature, humidity, precipitation, carbon dioxide levels, sunlight, etc.) and on the ground (canopy types, carbon dioxide taken in and released during photosynthesis, foliage produced, etc.). Called FLUXNET, the network of instrumented towers automatically records data every five minutes. These data are stored and later compared with satellite measurements, as well as model results.

It is the ongoing comparison of model results with real data that enables scientists to continue refining their models toward a real predictive capability. Already, climate models are showing promise that they will enable scientists to forecast certain changes months or even years ahead of occurrence.

Can models show us the future?
Why Build Computer Models?

FLUXNET towers are positioned all over the world - including this measuring site at Brasschaat, Belgium. (Photograph courtesy University of Antwerpen, Dept. of Biology, UIA, Wilrijk, Belgium Prof. R. Ceulemans)

Can models show us the future?

Will computer models enable scientists to accurately predict the future? Running says it is too early to answer that question, but some tantalizing results are emerging from current Earth-system modeling. For example, if the current climate models are correct in predicting higher temperatures coupled with greater rainfall in the next century, then the biospheric models predict a response of generally higher plant productivity and longer growing seasons. However, this seemingly positive prediction is tempered by the predictions that agricultural areas will gradually migrate to higher latitudes, leaving some current croplands too hot for production and severely disrupting local economies.

Running notes that his biospheric model results are directly coupled with climate model results. If higher temperatures occur without a corresponding increase in rainfall, then plant productivity is predicted to drop due to more frequent and widespread droughts.

Can we trust computer models?
Of course, models have limitations too. For instance, they can over-simplify the systems they are designed to simulate; thereby leading to erroneous results (Waring and Running 1998).

"If you were looking at a complex system like the biosphere and you tried to completely represent all the complexity of that system, you would drown in details," Running states. "Models are elegant simplifications of reality. There are many details that are left out because the models are generally focused on certain central tendencies of the ecosystem."

Therefore a model must be thoroughly tested in a wide range of conditions so that its strengths and weaknesses are well understood. Running also cautions that while they are important research tools, models cannot ever verify what the truth is (or will be). Only direct measurements can establish "scientific truth." But models can tell scientists where conventional understanding is wrong and encourage them to make the critical measurements that might not otherwise be made.

The terrestrial biosphere in the 21st century...

Using the same models that generated the trio of net primary productivity images on page 1, this set compares the model outputs based on predicted climate if atmospheric carbon dioxide doubled (710 parts per million). Scientists evaluate the accuracy of computer models both by comparing them to observations and comparing several models run with the same initial data. (Image courtesy of the University of Montana Numerical Terradynamic Simulation Group)

Running sees larger things in store for Earth scientists after the fall 1999 launch of NASA’s Earth Observing System flagship satellite–Terra. This new satellite carries a payload of sensors that will provide a new suite of data products, which Running feels will ultimately yield a new generation of models. With daily global measurements of land surface vegetative cover, forest fires and the amount of biomass burned, total leaf area and production rates of foliage, and incoming photosynthetically active (solar) radiation, Terra will provide direct measurements of many variables that today we can only model. Moreover, Terra will make measurements that apply to a wide range of atmospheric and oceanic disciplines, as well as land-based, so it addresses the Earth system as a whole.

"Finally, we will be able to answer questions like ‘Where is the missing terrestrial carbon sink?’ " Running surmises. "Terra’s measurements of the net primary production of green vegetation will help us quantify the terrestrial biosphere’s role in the global carbon cycle."

References

How to Build Better Climate Models

The Terra satellite, as it will appear after the solar panel and antenna deploy. (Image by Reto Stöckli, NASA GSFC Visualization Analysis Lab)