Maximum and Minimum Temperature Variations in a Perturbed Climate and Implications for Specialized Agriculture in the Great Lakes Region

Julie A. Winkler and Jeffrey A. Andresen
Michigan State University

Objectives: Impact studies primarily have dealt with projected changes in mean climatic states, which may poorly reflect the possible influence of a perturbed climate on natural ecosystems and human activities. Instead, an indication of the expected changes in the frequency of extremes, such as the occurrence of specified conditions and thresholds, is necessary. Thus, the challenge facing modelers and impact researchers is how to develop realistic climate scenarios at the requisite time and space scales.

Specialized agriculture (i.e., fruit and vegetable production) provides a particularly challenging system in which to evaluate the possible impact of climate change, as it is sensitive to temperature extremes. For example, deciduous fruit trees are vulnerable to cold damage during autumn before the tree is adequately hardened, during the dormant period when severe cold events may freeze flower buds and wood tissue, and during spring bloom when temperatures slightly below freezing may kill flower buds following the loss of cold hardiness. An adequate evaluation of the potential impact of a perturbed climate on fruit production requires, at a minimum, time series of daily maximum and minimum temperature.

The objectives of this study are to 1) evaluate the adequacy of daily simulations of screen temperature from a particular GCM, the Canadian Climate Center (CCC) model, for application in agricultural impact analyses at fine time and space scales, 2) develop statistically--derived transfer functions for "downscaling" temperature simulations from the CCC GCM to a local, daily scale for six locations in the fruit-growing areas of the Great Lakes region, and 3) analyze the transfer-function series for a perturbed climate in terms of thermal parameters relevant to specialized agriculture.

Products: 1 ) A detailed evaluation for the Great Lakes region of CCC model control (1xCO2) simulations of daily maximum and minimum temperature, sea-level pressure, and 500 mb geopotential height. 2) Evaluation and comparison of several regression-based transfer--function methodologies. 3) Transfer function--generated daily series of maximum and minimum temperature for control and perturbed (2xCO2) climates for six lake-influenced locations. 4) An assessment of the changes in the frequency or timing of critical temperature thresholds under a perturbed climate scenario. 5) Estimation of the possible changes in growing season heat accumulation for a perturbed climate.

Approach: Our first task was to evaluate whether CCC GCM daily temperature simulations can be directly utilized in impact studies. Simulations for nearby gridpoints were interpolated to a site with a high-quality observational record (Eau Claire, Michigan). The observed and simulated series were then compared using frequency distributions, daily and monthly descriptive statistics, and extreme value analysis.

The second task was to develop and evaluate transfer function methodologies. The methodologies utilized here are based on the National Weather Service's Perfect Prog (PP) and Model Output Systems (MOS), originally developed for short-range forecasting. Both the PP and MOS systems use regression techniques to formulate specification equations for individual locations which relate predictor variables, primarily free atmosphere variables, to predictands, typically surface-based variables. These techniques are based on the assumptions that free-atmosphere variables are forecasted more accurately than surface parameters and that regression equations implicitly incorporate local, site factors.

Applying the PP and MOS methodologies to GCM simulations first required that 1 ) the accuracy of the GCM simulations of the predictor variables be assessed, and 2) the impact on the transfer function-generated series of several user decisions such as function form and seasonal definition be evaluated. To accomplish the latter task, several variants on the regression-based methodology were explored. All transfer functions were calibrated using observations for the 1975-84 period and validated on the 1965-74 period before being applied to the CCC GCM simulations. The candidate predictor variables were limited to parameters derived from 500 mb geopotential height (Z500) and sea-level pressure (SLP).

The simplest of the transfer function methods we explored, which we have grouped under the heading "perfect prog", relates the predictor variables to the predictand (daily maximum or minimum temperature) via stepwise regression either for the entire year (PN), traditional or "fixed" seasons (PS), or for "floating seasons" (PD). The advantage of using floating seasons is that the beginning and ending dates of the periods are allowed to shift for a perturbed climate, when traditional seasonal definitions may be inappropriate. In order to account in part for any error in the GCM-simulated SLP and Z500 fields, transfer functions also were developed using standardized predictor variables (types ZN, ZS, and ZD). A third set of transfer functions were developed using the natural log of the predictand (LN, LS, and LD). Although the relationship between the predictors and predictand is highly linear for the 1975-84 calibration period, a comparison of the linear regressions against a more conservative logarithmic relationship allows us to partly assess the possible consequences of assuming a linear relationship outside the period for which the equations were developed (i.e., for a perturbed climate). Finally, transfer functions were developed using standardized predictor variables and the natural log of the predictand (XN, XS, and XD). The transfer function types were evaluated by comparing 1) daily and monthly statistics of the observed 1965-74 series and the transfer function-generated series for the validation period, 2) monthly statistics for the transfer function-generated 1 xCO2 series to observed values for 30 overlapping decades, and 3) daily mean and extreme temperatures for the lxCO2 period to 1950-1979 observed values.

Results to Date:

GCM1xCO2 Comparisons. The comparison of the maximum and minimum temperature series for the original CCC control simulation (GCM1xCO2) with observed series for Eau Claire indicated that the GCMlxCO2 series are not a satisfactory representation of present-day conditions. Although modeled mean annual maximum and minimum temperature are within 1.5°C of the observed values, several monthly means, particularly during spring, substantially differ from observed values. At the daily scale, the GCM1xCO2 daily means for both maximum and minimum correspond well with observed values from the end of July into December. During the remainder of the year, daily means for the two series deviate substantially, particularly for minimum temperature with the GCM--simulated daily mean persistently below the observed value in late winter and early spring and above the observed in summer and early fall. A striking characteristic of the GCM1xCO2 simulation is the unrealistic 'thresholds' during spring and autumn when both maximum and minimum temperature hover near freezing for extended periods and daily temperature range is extremely small. These "thresholds" also are evident in the CCC GCM2xCO2 series, although their timing and extent differs from the control period. As a result, GCM daily and monthly statistics, and also GCM seasonal and annual statistics, are questionable for locations where freezing temperatures are frequent. In particular, perturbed minus control comparisons for either point locations or latitudinal bands are potentially invalid. A comparison of daily means of CCC GCM control simulations of SLP and 500 mb height to observed values indicates that these variables are much more accurately simulated that screen temperature. Average daily SLP for the GCMlxCO2 series is smaller than observed from July through September, but otherwise the daily averages agree well. The daily means of Z500 for the GCM1xCO2 series and observed series agree well from September through March. During the remainder of the year, the GCM-simulated Z500 values are slightly too low.

Transfer function validation. Evaluation of the statistics for the validation period (1965-74) indicates that all the transfer function types adequately simulate the observed climate. Root mean square errors range from 2.9 to 3.7°C, comparable to error terms for short-range MOS and PP forecast systems which use a larger set of predictor variables (Andresen, unpublished data). In general, minimum temperature is better simulated by the transfer functions than maximum temperature. The error terms are smaller for those transfer function types with either fixed or floating seasons than those without seasons, particularly during March-May.

A comparison of monthly mean temperatures for the GCMlxCO2 series and the transfer function--generated series for the control climate to observed values for 30 overlapping decades (1946-84) reveals that the transfer functions much better represent the present climate than the GCM series, particularly during spring where the GCM average monthly temperatures are as much as 8°C colder than observed values. The largest errors for all transfer function types occur in spring, particularly for the methods that do not employ seasons. Also, the types with standardized predictor variables perform best. The methods with log transformed predictands perform similarly to the comparable linear versions.

Perturbed Scenarios. I n general, perturbed scenarios of maximum temperature generated using the transfer functions are warmer than the GCM2xCO2 series, particularly from November-May. On the other hand, the transfer function minimum temperature scenarios are colder than the GCM perturbed series. The greatest differences are found in January-February and June-September. An important finding of our analysis is that the perturbed series for the different transfer function types vary markedly, and impact researchers must carefully consider the type of transfer function when interpreting scenarios developed using statistical downscaling methods. Of the different transfer function methodologies, the PN, PS, LN, and LS perturbed scenarios are the most conservative, whereas the forecasted temperatures are warmest, particularly from March-May for the ZN, XN, ZN, and ZD scenarios. Only slight differences exist between the linear and logarithmic functions.

There is considerable agreement among the different transfer function types on the occurrence of first and last frost in a perturbed climate. The scenarios suggest the median first day of frost at Eau Claire will be delayed from the present by 25-35 days, and the last frost will occur 17-24 days earlier than present. Thus, the transfer function scenarios suggest that the growing season length in the fruit production region along Lake Michigan will be considerably extended in a 2xCO2 environment. The more extreme scenarios (ZD, XD, ZN, XN) suggest the number of days per growing season >35°C, a general temperature threshold above which plant photosynthesis and other processes begin to slow, will increase by more than 20 days, whereas the more conservative scenarios suggest an increase of fewer than 10 days >35°C in a perturbed climate.

Manuscripts in Preparation: The simulation of daily time series from GCM output. Part 1: Comparison of model data with observations. J.P. Palutikof, J.A. Winkler, C.M. Goodess. and J.A. Andresen. The simulation of daily time series from GCM output. Part 2: Development of local scenarios for maximum and minimum temperature using statistical transfer functions. J.A. Winkler, J.P. Palutikof, J.A. Andresen, and C.M. Goodess.

Conference Preprints and Proceedings: Local climate change scenarios of maximum and minimum temperature for Michigan. J.A. Andresen, J.A. Winkler, J.P. Palutikof, and C.M. Goodess. Proceedings, First Symposium on Michigan Weather and Forecasting, in press.

The simulation of daily time series from GCM output. Part 1: Comparison of model data with observations J.P. Palutikof, J.A. Winkler, C.M. Goodess, and J.A. Andresen. Preprints, Sixth Symposium on Global Change Studies, January, 1995, Dallas, Texas, pp. 45-48.

The simulation of daily time series from GCM output. Part 2: Development of local scenarios for maximum and minimum temperature using statistical transfer functions. J.A. Winkler, J.P. Palutikof, J.A. Andresen, and C.M. Goodess. Preprints, Sixth Symposium on Global Change Studies, January 1995, Dallas, Texas, pp. 93-98.

Presentations: Development of local scenarios of maximum and minimum temperature using statistical transfer functions. J.A. Andresen and J.A. Winkler, First Symposium on Michigan Weather and Forecasting, March 1995, Ann Arbor, Michigan.

Local climate change scenarios of daily temperature for Michigan: Development and Application. J.A. Winkler, J.A. Andresen, J.P. Palutikof, and C.M. Goodess. 91st Annual Meeting, Association of American Geographers, March 1995, Chicago, Illinois.

The simulation of daily time series from GCM output. Part 1: Comparison of model data with observations. J.P. Palutikof, J.A. Winkler, C.M. Goodess, and J.A. Andresen. Sixth Symposium on Global Change Studies, January, 1995, Dallas, Texas.

The simulation of daily time series from GCM output. Part 2: Development of local scenarios for maximum and minimum temperature using statistical transfer functions. J.A. Winkler, J.P. Palutikof, J .A. Andresen, and C. M. Goodess. Sixth Symposium on Global Change Studies, January, 1995, Dallas, Texas.

The construction of regional climate change scenarios using a transfer function methodology. J.A. Winkler. Department of Geology Seminar Series, Michigan State University, November 1994.

Student Participation: Greg Bierly 50%, Chi-Pi Chou 25%. Eleanor Waller 25%.