Top Navigation Bar

ASAE Journal Article

Evaluation of Winter Freeze Damage Risk to Apple Trees in Global Warming Projections

M. Baraer, C. A. Madramootoo, B. B. Mehdi


Published in Transactions of the ASABE Vol. 53(5): 1387-1397 ( Copyright 2010 American Society of Agricultural and Biological Engineers ).

Submitted for review in March 2009 as manuscript number SW 7608; approved for publication by the Soil & Water Division of ASABE in July 2010.

The authors are Michel Baraer, Doctoral Candidate, Brace Center for Water Resources Management, Faculty of Agricultural and Environmental Sciences; Chandra A. Madramootoo, ASABE Fellow, Dean, Faculty of Agricultural and Environmental Sciences; and Bano B. Mehdi, Research Associate, Brace Center for Water Resources Management, Faculty of Agricultural and Environmental Sciences, McGill University, Montreal, Québec, Canada. Corresponding author: Michel Baraer, Brace Center for Water Resources Management, Faculty of Agricultural and Environmental Sciences, McGill University, 21111 Lakeshore Road, Montreal, Québec, Canada H9X 3V9; phone: 514-398-7833; fax: 514-398-7767; e-mail: Michel.Baraer@mail.mcgill.ca.


Abstract. Winter freeze damage affects fruit production regularly in the northern part of North America. This situation, which is related to climatic conditions, financially affects fruit producers and limits the affected areas to the use of cultivars that are freeze-resistant but do not always yield a sufficient market return. The purpose of this study is to conduct an experiment with a newly developed numerical model (W5L+) and its associated snow cover module to evaluate the effects of the projected climatic change on the risk of winter freeze damage to apple trees. The model W5L+ quantifies the risk of freeze damage occurrence at defined locations based on local meteorological records or projections. Risk quantification is achieved by screening daily meteorological time series with pre-identified parameters that are known to be proxies for conditions that result in freeze-damage. The model was parameterized using historical meteorological records from apple orchards in Farnham, southern Québec, and descriptions of regional winter freeze damaging events that were recorded between 1920 and 2005. In 82% of the years studied, the model was able to identify correctly the order of magnitude of the recorded freeze events. During the same period, results suggest that extremely low temperatures and prolonged periods of low temperatures were responsible for the majority of damaging events. When used with climatic projections downscaled from a global climate model (GCM), the model predicted a decrease in freeze risk for apple trees at the Farnham orchards in the next 60 years. This trend is due to a decrease in extreme cold events as well as in prolonged periods of low temperature. The present study demonstrates the potential of the W5L+ modeling approach in studying the impact of climate change on the occurrence of damaging freezes. However, the predictions need to be verified by using the model with a large range of agro-climatic conditions and climate projections.

Keywords. Apple trees, Climate change, Computer models, Freeze damage, Snow cover.

Freeze events in apple orchards can damage trees and buds, limiting apple production for several years. In 1981, 236,000 apple trees were killed in Québec due to exceptionally unfavorable winter conditions (Charette and Krueger, 1992). Important killing freeze events that have a regional impact on apple trees in Québec are recorded approximately once every ten years. This figure does not take into consideration minor or more localized freeze events that affect only a limited number of producers. These small-scale events occur more frequently, but are not systematically recorded. Winter damage to fruit trees generally occurs in areas where the minimum temperatures are far below 0°C.

Progress has recently been made in identifying cultural factors associated with winter injury to apple trees (Khanizadeh, 2007; Khanizadeh et al., 2009). Potential mitigation measures arising from this research should help producers reduce the losses related to extreme winter conditions. On a larger scale, the economic consequences of freeze events in orchards remain important. These consequences are of two forms: first, loss of trees limits regional production capacity and affects revenues for several years, and second, freeze risk limits the geographical distribution of freeze-sensitive varieties. The latter is the case in eastern Canada, where some species, such as apricots or peaches, can be produced only in the Niagara region and along the northern shore of Lake Erie due to their vulnerability to cold temperatures (Rochette et al., 2004). In Québec, freeze occurrences inhibit further development of apple production. Several commercially viable cultivars that are well received by consumers, such as Gala, are not sufficiently hardy to support profitable commercial operations.

With an increase in global temperatures predicted for the coming decades (IPCC, 2007), the frequency and distribution of freeze events may change. Describing this potential change is made difficult by the complexity of the mechanisms involved. Unlike the damage caused by a spring frost, it is difficult to attribute winter damage at a particular stage of growth to a given temperature (Eccel et al., 2009). Winter freeze events are related to the inability of trees to adapt their resistance to low temperatures during particular meteorological events. This adaptation function, called freeze hardiness, is an extremely complex phenomenon. It is affected by temperature, length of day, and conditions of the tree such as maturity, water content, nutritional stage, physiological age, and dormancy status (Palonen and Buszard, 1997). Consequently, predictions of the impact of climate change on the occurrence of freeze events and intensity are limited in number, making it difficult to estimate future tree-based production (Jonsson et al., 2004). Studies undertaken to date are of two general types: studies that target cold hardiness modeling as a function of environmental parameters, and studies that are based on agro-climatic parameters used as indicators for potential freeze conditions. For example, Kobayashi (1983) developed a cold hardiness, non-mechanistic model but concluded that the model results may not be accurately reproduced under field conditions. Leinonen (1996) drew similar conclusions, underlining the fact that one of the limits to the use of models based on biological processes is the large amount of experimental work and the number of long-term field observations that are needed to replicate sites and cultivar characteristics.

To circumvent this modeling problem, studies of the impact of climate change on freeze events and intensity undertaken so far have been based mainly on analyses of isolated agro-climatic parameters that are recognized as key factors behind freeze damage. For example, Quamme et al. (2010) associated freeze damage to fruit trees in the Okanagan Valley of British Columbia, Canada, with arctic outflow events. Prediction of climate change impacts was then achieved by studying the trend of these climatic events and by hypothesizing that the trend will remain unchanged in the future. Jonsson et al. (2004) developed climatic indexes for some of the potentially damaging conditions for trees. A similar approach was used by Rochette et al. (2004) in studying the effects of climate change on winter damage to fruit trees in eastern Canada. In their study, five indexes were calculated from measured and projected temperatures for three different time periods: 1961-1990, 2010-2039, and 2040-2069. This pragmatic approach enabled the authors to draw hypotheses about freeze damage trends based on projected variations in agro-climatic parameters. However, their approach in itself did not allow for evaluation of the combined effects of changes in agro-climatic conditions. Such combinations are of major importance in the freeze processes (Palonen and Buszard, 1997).

In this study, we combine the modeling approach and the agro-climatic parameters approach to describe the impact of climate change on the risk of winter freeze damage to apple trees. The resulting model quantifies annual freeze risk by means of a freeze index that is calculated from both meteorological records and climatic projections for the Farnham meteorological station in Québec, Canada.

Study Area

The present study focuses on apple trees ( Malus domestica Borkh) and specifically the McIntosh cultivar. Farnham, the study site, is situated in Montérégie, Québec, Canada (fig. 1). With its 3530 ha of production area, Montérégie is the province's largest apple-producing region (Institut de la statistique Québec, 2007). Farnham meteorological data were recorded from 1920 to 2005, except for the 1941 to 1952 period. The parameters that were measured include the minimum, maximum, and average daily temperatures and the daily liquid, solid, and total precipitation. Recorded data for snow on the ground are available for 1981 to 2005. Times series are complete for almost all years. In the occasional instance for which there were no data, an extrapolation of nearby station records completed the time series. Extrapolated values were obtained by calculating the geometrical mean of data from the Rougemont, Sutton, and Philipsburg stations, situated 12, 35, and 29 km, respectively, from Farnham. Records from the Farnham meteorological station provide a good indication of winter conditions in Montérégie. During the 1972-2000 period, January was the coldest month of the year, with a mean temperature of -10°C. In total, four months of the year have mean temperatures that are below 0°C. The average lowest minimum temperature of the year during the same period was -31°C. The lowest absolute temperature recorded between 1972 and 2000 was -42°C, measured in January 1994. The mean annual precipitation was 1 077 mm, 17% of which reached the ground as snow.

SW7608_files/image1.gif

Figure 1. Map of Montérégie, Québec, Canada. Farnham, the study site, is marked by the black circle. The weather stations that were considered for data extrapolation are marked by triangles.

Method

Climatic Projections

General circulation models (GCMs) are the best available tools to use in estimating future global climate changes resulting from the increasing concentration of greenhouse gases in the atmosphere (Busuioc et al., 2001). Because of the trade-off between physical detail and computational power, GCMs can provide only a crude spatial resolution of global grids (Russo and Zack, 1997). This spatial resolution is too coarse to resolve a regional-scale effect or to use directly in local impact studies. Downscaling techniques provide an alternative means to improve regional or local estimates of variables from GCM outputs (Hessami et al., 2008). In the present study, we use the stochastic weather generator LARS-WG to downscale Canadian global circulation model CGCM1 daily data produced under the middle range and widely used IPCC scenario IS92a. LARS-WG has been shown to perform well in simulating different weather statistics, including the climatic extremes relevant to agriculture (Semenov and Barrow, 1997). LARS-WG applies global climate model-derived changes in precipitation or temperature to generate synthetic weather time series from observed weather data. Times series that have been produced for the same period have the same statistical distribution, and variations in yearly predictions within the period are purely random (Baraer and Madramootoo, 2007; Semenov and Barrow, 2002).

LARS-WG was calibrated using the 1972-1985 records from the Farnham meteorological station. The accuracy of the calibration was verified using the LARS-WG Qtest option, which carries out a statistical comparison ( ? 2 , t-tests, and F-tests ) of 300 synthetic time series of 13 years with the parameters derived from observed weather data (Semenov and Barrow, 2002). The downscaling setup was validated by comparing the 1985-2005 records from the Farnham meteorological station to the time series generated by LARS-WG. Finally, synthetic weather time series were based on the CGCM1 projections for three different time periods: 1972-2000, 2010-2039, and 2040-2069.

Freeze Events Records

Modeling freeze events requires historical records of freeze observations for calibration and validation of the model. In order to create a list of such observations, available documents on regional apple production conditions were systematically reviewed and information about freeze events in the Montérégie region was recorded. The screening covered the 1920-2005 period to match the time series from the Farnham meteorological station. Each year for which there was usable meteorological data was then described by an observed freeze index. Ranging from 0 to 4, the index indicates the strength and intensity of freeze events reported in the literature. The criteria for index attribution are given in table 1.

Table 1. Definitions of yearly freeze indexes based on historical descriptions.

Index

Definition

0

Existing statement(s) describing a year without winter damages.

1

Description(s) of a possible or very localized light damaging freeze or year with no freeze event description.

2

Description(s) of a localized freeze with minor consequences.

3

Description(s) of a localized freeze event with significant consequences locally or a regional freeze event with measurable, but not major, consequences.

4

Several descriptions of a major freeze with significant region-wide loss.

Several different sources were used to accumulate the required observations. The annual reports of the Pomological Society of Québec (QSP, 1870-1953) were the most important information source for the first half of the 20th century. More recent events are described in the literature (Charette and Krueger, 1992; Coleman 1991; Chouinard, 1997), in government publications (Audette and Ditcham, 2007; Bergeron et al., 1997), and in other reports issued by organizations, such as Agri-reseau (2007) and Recupom (2007). The years 1941 to 1952 were excluded from freeze observation due to insufficient records.

Snow Cover Module

A significant amount of snow on the ground can protect roots from extremely low air temperatures and help to prevent damage that such conditions may cause (Coleman, 1991). Therefore, snow cover needs to be considered in freeze risks assessments, although it cannot be easily downscaled from a GCM. In the present case, we used a specific module called Sn 2 to generate snow cover projections based on climatic parameters downsized from the GCM.

Sn 2 calculates the depth of snow on the ground ( S n , cm) for day n using the maximum daily temperature ( TM n , °C), the daily amount of snowfall ( Ps n , cm), and the snow cover of the previous day ( S n- 1 ):

SW7608_files\eqn\eqn1.gif (1)

where parameters a , b , and c are site-specific factors that are fixed, using trial and error, by module calibration against measured values. In the present case, we used the Farnham snow-on-the-ground time series for 1981 to 1991 for calibration and validation. The performance of Sn 2 was measured by calculating the root mean square error (RMSE) as well as by counting the number of days for which the module misreported the presence or absence of snow on the ground. The RMSE is an indicator of how well the model reproduces measures of snow cover depth. The second indicator, called "no match," targets the way W5L+ considers the snow-on-the-ground parameter. To compute the freeze risk index, W5L+ does not use the snow depth parameter directly. Instead, it converts snow depth to the presence or absence of snow on the ground. The "no match" indicator sums up all cases where Sn 2 predicted snow on the ground that was not recorded in the historical data, and vice versa. Both indicators were calculated for winter days only and not for complete years. The performance of Sn2 compared with published models was evaluated by applying the snow cover model used by Singh and Bryant (2006) to southern Québec. We name this model "Scott" because it is adapted from the method used by Scott et al. (2002) in their study of the vulnerability of winter recreation to climate change in Ontario. We calibrated and validated the Scott model, described in equation 2, for exactly the same conditions used for Sn 2 :

SW7608_files\eqn\eqn2.gif (2)

where Pl represents liquid precipitation (mm), SW7608_files\eqn\eqn3.gif represents average daily temperature (°C), and parameters d and k are considered site-specific. Their values are calculated by calibration.

W5L+ Model Overview

W5L+ is based on pattern recognition. The model concept is to screen meteorological time series for conditions that have been reported as involving winter tree-damaging events. A distinction is made between conditions that are direct causes of damage to trees (damaging condition) and conditions that make trees more vulnerable to harmful conditions but are not directly responsible for damage (amplifying factor). Amplifying factors include conditions that trigger changes in tree hardiness, including acclimatization and unacclimatization processes, as well as environmental factors that protect trees from damaging conditions. For example, an extremely low temperature event can cause intracellular freezing that, by destroying cell membranes, would lead to necrosis (Linden, 2002). Therefore, the extreme temperature event is classified as a "damaging condition". As described previously, snow cover can protect a tree from punctual extreme temperatures by creating a layer of insulation between the air and the soil. Therefore, lack of snow can be considered as an amplifying factor that makes the trees vulnerable to temperatures that would not have caused harm if there had been snow. The model W5L+ was first used with Farnham meteorological data recorded between 1920 and 2005 to study past freeze events. As a second step, synthetic meteorological data generated for the years 1972 to 2069 were used to produce freeze risks projections.

Damaging Conditions

The four damaging conditions considered in W5L+ were selected on the basis of what is reported in the literature (table 2). For each condition, a corresponding set of measurable parameters is used in W5L+ to detect the condition in meteorological records or projections. The first condition (extremely low temperature) corresponds to single events of extremely low temperature (Cline et al., 2005). Below a given temperature, ice may form in the tree's cells, leading to cellular necrosis (Linden, 2002). Because the damage thresholds vary as winter progresses (Burke et al., 1976), four separate time periods are considered when calculating the corresponding parameters. The four periods represent four possible hardening phases. The second condition (periods of low temperature) covers cold episodes during which the minimum temperatures do not necessarily reach the extremes described above, although low temperatures are recorded on several days, potentially causing cellular desiccation (Charette and Krueger, 1992). W5L+ screens for this condition by computing the number of days during which the minimum or average temperatures is below a given value for a fixed period. The third damaging condition is temperature drop. In winter, trees need time to adapt from standard temperatures to very cold events. An abrupt drop in temperature can cause damage by not allowing the tree sufficient time to adapt (Coleman, 1991). Related parameters used in W5L+ cover single occurrence as well as repeated temperature drop events. Finally, the fourth condition (mild episodes followed by low temperatures) includes conditions that may trigger tree dehardening once dormancy is over. Such dehardening takes place during the second part of the winter, when mild episodes of several days in length are followed by temperatures that are well below 0°C (Kaukovirta and Syri, 1985). W5L+ covers this condition type by screening for days that have a positive average temperature in advance of a low minimum temperature.

Table 2. Classified damaging conditions and corresponding parameters used in the W5L+ model. T min , T max , and T avg represent the daily minimum, maximum, and average temperatures (°C), respectively.

Damaging Condition

Corresponding Parameter

References

I1: Extremely low temperature

I1a: T min between November 15 and December 15

Charette and Krueger, 1992;

I1b: T min between December 16 and January 1

Linden, 2002; Coleman, 1991;

I1c: T min between January 1 and February 28 or 29

Burke et al., 1976

I1d: T min between March 1 and March 31

I2: Periods of low temperature

I2a: No. of days with T avg < -18°C in a moving 21-day period

Linden, 2001; Charette and

I2b: No. of days with T min < -30°C in a moving 21-day period

Krueger, 1992; Hill, 1941

I3: Temperature drop

I3a: T max - T min (on the same day) when T min < -18°C

Charette and Krueger, 1992;

I3b: T max - T min (on two consecutive days) when T min < -18°C

Coleman, 1991; Kaukovirta

I3c: No. of days per winter with T max > 0 ° C and T min < -20°C

and Syri, 1995; Linden, 2001

I4: Mild episodes followed by

I4a: No. of consecutive days with T avg > 0°C in a 20-day period before T min < -28°C

Charette and Krueger, 1992;

low temperatures

I4b: No. of consecutive days with T avg > 0°C in a 20-day period before T min < -32°C

Linden, 2002; Coleman, 1991;

I4c: No. of days with T avg > 0°C in a 30-day period before T min < -26°C

Wees, 2001; Pool et al., 2004

I4d: No. of days with T avg > 0°C in a 20-day period before T min < -26°C

I4e: No. of consecutive days with T avg > 0°C in a 5-day period before T min < -26°C

In using W5L+, the risk associated with a damaging condition-related parameter is determined by comparing its computed value to pre-established thresholds. Four different thresholds are assigned to each parameter, resulting in a risk index that ranges from 0 to 4, with 0 representing no risk of damage and 4 representing a high risk of freeze damage. Thresholds for the extremely low temperature condition (I1) apply only to specific days of the year and not for a longer period. W5L+ applies linear extrapolation between two given dates to calculate a threshold for each day of the period. Preliminary threshold sets are derived from frost events described in the literature (table 2). As some of the preliminary thresholds or damaging conditions taken from the literature are given for trees other than Malus domestica Borkh, final values are set by a model calibration stage, whereas trial and error is used to obtain the best possible match between the reported freeze events intensities and the synthetic freeze events.

Amplifying Factors

The incorporation of amplifying factors in the model was based on what has been reported in the literature. At least one measurable parameter is associated with each factor. These parameters, calculated from meteorological time series, are compared to a unique threshold, leading to a presence/absence binary system. Table 3 lists the factors selected in this study. It is suspected that two of these factors (high precipitation in the fall and mild fall) are responsible for delays in tree hardening. Such delays make trees more susceptible to freeze damage once there are winter conditions (Linden, 2002). In W5L+, these two factors are sought by comparing the September to November daily average precipitation (mm) and the October 15 to November 30 daily average temperature (°C), respectively, to fixed threshold values. In both cases, the threshold is defined as the 90th percentile of the same parameter calculated for the years 1980 to 2005 (data for 1997 and 1998 are missing).

Table 3. Classified amplifying factor and corresponding parameters used in the W5L+ model. T avg represents the average daily temperatures expressed in degrees centigrade.

Amplifying Factor

Corresponding Parameter

Reference

High precipitation in the fall

F1: September 1st to November 30 average precipitations (mm)

Linden, 2002

Mild fall

F2: October 15 to November 30 average temperature (°C)

Beck et al, 2004

Lack of snow cover

F3: Snow on ground (cm)

Coleman, 1991

Dry summer

F4: June 1st to August 30 average precipitation (mm)

Charette and Krueger, 1992

Early fall low temperatures

F5a: Minimum temperature in September (°C)

Charette and Krueger, 1992

F5b: Minimum temperature in October (°C)

Late harvest

F6: Growing degree-days for T avg > 5°C calculated from March 1 to Sept. 30

Pool et al, 2004

The previously described lack of snow cover is identified by comparing snow on the ground (cm) to a calibrated threshold. By not giving the tree an opportunity to mature, dry summer conditions make the tree more susceptible to freeze damages (Charette and Krueger, 1992). A search for the condition of wet fall is conducted by comparing the season's average precipitation (June to August, in this case) to the 10th percentile of the same parameter calculated for the 1980-2005 period. Two parameters are used to cover the condition of early fall low temperatures. By shortening the nutrient transfer period, early freezing temperatures decrease a tree's capacity to adapt to very low temperatures. The minimum daily temperature is used here as a measurable indicator of early low temperatures. Two different periods, September and October, are considered for this indicator. By giving insufficient time to the tree to accumulate nutrients, a late harvest produces effects close to those generated by an early fall low temperature (Pool et al., 2004). As it was not possible to use past harvest dates in the W5L+, an evaluation of the annual time at which fruits mature is provided by calculating the growing-degree-days from March 1 to September 30. Growing degree-days (GDD) are calculated by adding, for the period under consideration, the daily positive differences between the day's mean temperature and 5°C. Negative differences are ignored. A late harvest is assumed when the calculated GDD value is lower than a specific threshold value. As for damaging conditions, amplifying factor thresholds were set on the basis of the literature and adjusted during model calibration.

In W5L+, when an amplifying factor is detected, thresholds that apply to damaging parameters are shifted (by a given factor) to make the damage more likely to happen. For example, if a mild fall is identified, the thresholds that are applied to the minimum daily temperature (I1 parameters), are increased. Amplifying factors are cumulative. This means that a shift rate is used for each detected factor. The rate of shift for thresholds is the same as for all damaging condition parameters.

Event freeze indexes are not cumulative. The yearly freeze index is obtained by taking the absolute highest event index of the year. For example, if two events of level 1 are obtained for one winter, the yearly freeze index is still 1.

Results and Discussion

Climatic Projections

A LARS-WG integrated Qtest conducted on the 1972-1985 calibration output was used to assess the success of the calibration of the downscaling tool. The Qtest results showed that there were no significant differences ( a = 0.05) between the observed and the synthetic mean results for the key parameters used in the present study ( T min , T avg , and precipitation) and that the resulting setup was ready for validation tests. The validation phase that was performed on the 1985-2005 Farnham meteorological records compared LARS-WG synthetic meteorological time series to real data (fig. 2). On a yearly basis, the differences between synthetic and observed values are small for both the maximum temperature and minimum temperature. Figure 2 shows an almost perfect fit on a monthly basis, except for the last two months of the year. Differences between values observed in November and December range from 0.9°C to 1.6°C. These validation results show both the strength and limits of the methods that we applied. For example, in December, the difference between the GCM and the historical data from Farnham was more than 10°C for a similar time window. With a difference of 1.6°C, the worst of the downscaled synthetic data are much closer to the observations than the GCM. However, a difference of 1.5°C in the monthly average is still important. Improving these meteorological projections would definitively constitute an improvement in the W5L+ model parameterization and evaluation.

SW7608_files/image5.gif

Figure 2. LARS-WG validation results (1985-2005): T min is the monthly average minimum temperature (°C), and T max is the monthly average maximum (°C).

The synthetic weather time series produced for use in this study were based on the CGCM1 output for three different time periods: 1972-2000, 2010-2039, and 2040-2069. The results for the monthly average minimum temperature are given in table 4. The projected increase in annual mean minimum temperature reaches 2.8°C between 1972-2000 and 2040-2069 at the Farnham meteorological station. This is more than double the increase in annual minimum temperature between the 1972-2000 and 2010-2039 periods, suggesting that the rate of temperature increase will not be constant during the next 60 years, but will gradually increase.

Table 4. Reproduced or projected monthly average minimum temperatures for the Farnham station.

T min (°C)

1972-2000

2010-2039

2040-2069

January

-16.4

-13.6

-11

February

-14.3

-11.5

-7.6

March

-7.2

-5.8

-3.1

April

1.2

1.2

2.1

May

7.4

7.4

8.2

June

12.2

12.9

14.4

July

14.7

15.2

16.7

August

12.9

14.1

15

September

8.2

10.2

11.5

October

3.2

4.7

6.3

November

-2.5

-2.3

-1.6

December

-10.8

-9.6

-8.8

Annual

0.72

1.91

3.51

Changes in the average minimum temperature are not evenly distributed throughout the year. It is predicted that the first three months of the year will exhibit a much more pronounced increase in temperature than the rest of the year. These months are three of the four coldest months of the year. This scenario predicts, therefore, that warming will be more pronounced in winter than during the rest of the year. In turn, this means that future yearly minimum temperature ranges will be smaller than those of the past 30 years.

Freeze Event Records

A review of annual freeze indexes based on historical records appears in table 5. Of the 69 years that were evaluated, ten had a freeze index of 4, and five had a freeze index of 3. Indices 3 and 4 correspond to years with important recorded freeze events, so these figures indicate that an important event happened on average every five years. This is slightly more frequent than the one major event every ten years calculated by Charette and Krueger (1992). One of the particularities of the table 5 time series is that two consecutive years of high freeze index are reported five times. This is the same number as reported for isolated high index years. There is only a very low probability that the high frequency is of random origin. Instead, the high frequency suggests that the two consecutive events are not independent of each other. The relationship between the two consecutive events could be physiological or circumstantial. In the first case, the first events would be an important amplifying factor for the conditions of successive winters. In the second case, tree damage from the first event could be detected and/or reported with a delay greater than six months, making freeze records from the second winter higher than they should be.

Table 5. Evaluated yearly freeze index based on historical descriptions. Winters are identified by the two last digits of the year in which the winter begins. Missing winters are due to missing meteorological records for Farnham or years that have conflicting information on freeze occurrence.

Winter

04

03

02

01

00

99

96

95

94

93

92

91

90

89

88

87

86

85

84

83

82

81

80

Index

0

2

2

0

1

2

1

1

1

4

4

1

1

3

1

1

4

4

1

1

1

3

4

Winter

79

78

77

76

75

74

72

71

70

69

68

67

66

65

64

63

62

61

60

59

58

57

56

Index

1

1

1

1

4

1

1

1

1

1

1

4

1

1

1

1

1

1

1

1

1

1

4

Winter

55

54

53

52

39

38

37

36

35

34

33

32

31

30

29

28

27

25

24

23

22

21

20

Index

1

3

1

1

0

0

0

0

0

3

4

0

0

0

0

0

0

1

4

3

1

1

0

Snow Cover

The best fit between observed and predicted snow cover in Farnham was obtained by using values of a = 0.104, b = 0.37, and c = 1.86 in equation 1. In this setup, Sn 2 gives an RMSE (winter period only) of 6.22 cm for the 1981-1991 period. Figure 3, which shows detailed synthetic values for the entire period that was studied, illustrates how well the model captures interannual differences in the shapes of the snow cover evolution. All years in which deep and long-lasting snow cover was observed, such as 1982, 1986, and 1987, are modeled by Sn 2 with similar attributes. The same modeling performance is observed for years that had a limited amount of snow on the ground, such as 1983 and 1988. Cases where measured snow on the ground was not reproduced by Sn2 occurred infrequently.

SW7608_files/image6.jpg

Figure 3. Sn 2 model results compared to the measured snow on the ground in Farnham for 1981-1991.

The main weakness of Sn 2 is probably in its ability to reproduce the height of peak events. This is the case for 1983 and 1989, where time-limited, snow-on-the-ground events were reproduced by Sn 2 , but not to the correct level. However, the way in which W5L+ integrates the snow-on-the-ground parameter ensures that the present study is not affected by this limitation.

Calibrated under exactly the same conditions as Sn 2 , the reference model (Scott model) with K = 0.21 and d = 0.34 produced an overall winter RMSE of 7.13 cm. This error measure is slightly greater than the RMSE produced by Sn 2 . Details of the differences between the performances of Sn 2 and Scott are given in figure 4.

SW7608_files/image7.gif

Figure 4. Yearly Sn 2 model errors as a function of the Scott model errors: ? RMSE corresponds to the Sn 2 model RMSE minus the Scott model RMSE, and ? no match is the difference between the Sn 2 annual no match and the Scott model no match.

The differences in RMSE and in "no match" are similar in sign for all years, although not in amplitude. This is the case, for example, in 1987, the year with the highest difference in RMSE, but with a "no match" indicator that shows no great differences in model performance. The opposite tendency is observed for 1981, 1982, and 1988, where almost no difference in RMSE is measured but where "no match" indicates a relatively good performance of the Sn 2 model. For 11 years of snow-on-the-ground data, the Scott performance indicators were better than the Sn 2 indicators only once, in 1989. Overall, figure 4 indicates that Sn2 is relatively well adapted to W5L+ requirements.

W5L+ Model Calibration and Verification

The entire set of historical data was used to calibrate the W5L+ model. Original thresholds established on the basis of event descriptions were adjusted by trial and error to reach an optimal fit between the calculated yearly freeze index and the evaluated index. The goal is for the model to identify those years that have reported freeze damage without predicting freeze hazards for years for which there are no damage records. Whether W5L+ performance achieves that objective is determined by calculating the projection error. This error is the difference between the modeled yearly freeze index and the index evaluated from historical records. The results of this evaluation appear in figure 5.

SW7608_files/image8.gif

Figure 5. Distribution of the projection error at the W5L+ verification stage. A negative error corresponds to an under-reproduced observed yearly index, while a positive error corresponds to an overestimate of what was observed.

The great majority (86%) of errors are within a range of -1 to +1. In view of the fact that evaluating a yearly index involves a certain degree of uncertainty, an error of ± 1 can be considered acceptable. This means that W5L+ was able to reproduce the magnitude of the freeze risk associated with a given year with reasonable accuracy. In 12% of the years studied, the absolute error reached 2.0. In the corresponding cases, W5L+ did not provide a misleading risk projection but also did not capture the amplitude of the risk. The projection made a mistake twice: once in 1986 when reported serious damage to trees was not detected by the model, and once in 1929 when W5L+ projected an index of 4 for a year with written statements of no freeze damage. Years with a negative related error are the most numerous. This is probably due to the rules that we applied to the years for which the literature had no observations. In many cases (13), the W5L+ generated index was 0 for these years. Overall, the majority of the year was fairly well modeled, and W5L+ showed that it was capable of detecting tendencies of a risk level that is associated with freeze damages. However, because of the two years of misleading projections, a year-by-year analysis of results is not recommended.

The final threshold values applied to damaging conditions and amplifying factors are presented in table 6. Fixed through a model calibration process, the thresholds that have been crossed at least once during that process are highlighted. The W5L+ setup shows that two damaging conditions (extremely low temperature and periods of low temperature ) are the conditions for which threshold ranges have been crossed most frequently historically. This confirms that these two conditions played a fundamental role in freeze damage occurrence in Farnham from 1920-2005. Unlike the parameters associated with these two "active" conditions, none of the parameters associated with mild episodes followed by low temperatures had thresholds 3 or 4 crossed at least once.

Table 6. Thresholds (Th1 to Th4) applied to classified damaging conditions and amplifying factors after calibration of W5L+. For example, a parameter passing Th1 but not passing Th2 leads to an event index of 1 and so on. Precipitation is expressed as liquid equivalent (mm). T min and T max represent minimum and maximum daily temperatures (°C), respectively. [a]

Damaging Condition

Th1

Th2

Th3

Th4

Extremely low temperature

I1a: T min

-24

-26

-31

-32

I1b: T min

-28

-29

-31

-32

I1c: T min

-37

-39.5

-40.5

-41.5

I1d: T min

-29

-32

-34

-36

Periods of low temperature

I2a: No. of days

3

4

5

6

I2b: No. of days

3

4

5

7

Temperature drop

I3a: T max - T min

37

38

39

40

I3b: T max - T min

35

36

39

40

I3c: No. of days

7

8

9

10

Mild episodes followed by low temp.

I4a: No. of consecutive days

4

5

6

7

I4b: No. of consecutive days

3

4

6

7

I4c: No. of days

10

12

14

16

I4d: No. of days

5

6

8

9

I4e: No. of consecutive days

--

3

4

5

High precipitation in the fall

Th

F1: Precipitation (mm)

4.1

Mild fall

F2: Temperature (°C)

8.4

Lack of snow cover

F3: Snow depth (cm)

4

Dry summer

F4: Precipitation (mm)

2.8

Early fall low temperatures

F5a: Temperature(°C)

-4

F5b: Temperature (°C)

-12

Late harvest

F6: Growing degree-days (°C·d)

1600

[a] Numbers in bold italics are thresholds that have been crossed at least once between 1920 and 2005 with the historical meteorological records from Farnham.

This indicates that W5L+ makes use of mild episodes followed by low temperatures in its freeze damage risk prediction, but does not recognize this pattern as being directly responsible for one of the most significant freeze events recorded during the 69 years considered in the study. Table 6 shows also that the variation in time of the extremely low temperature thresholds is well supported by the calibration results. Thresholds increase gradually from late fall to early winter, reach a peak during the two coldest months of the year, and then decrease gradually up to the beginning of March. However, the fact that we observe a significant difference between I1a and I1b for Th1 and Th2, but no differences for Th3 and Th4, suggests that the way in which those two parameters were established can be improved.

For the periods of low temperature, the first parameter (I2a) is daily average temperature-based, whereas the second parameter (I2b) uses the daily minimum temperature (table 2). In the W5L+ calibration period, I2b thresholds were used more often and at a higher level than I2a thresholds. This suggests that damages related to periods of low temperature are more driven by minimum temperature than dependent on average temperatures. The same observation applies to temperature drop; parameters I3a and I3b generated higher indexes values than parameter I3c. This supports the possibility that effects of temperature drops are event-related and not cumulative.

W5L+ Application: Past Freeze Risk Reproduction

Yearly freeze indexes generated by W5L+ based on the 1920-2005 meteorological records from Farnham appear in figure 6. To facilitate a visual trend analysis, the yearly indexes are grouped in three different periods: 1920-1950, 1950-1980, and 1980-2005. Figure 6 shows that the index is relatively stable during the period studied. Due to the uncertainty associated with the index evaluation and calculation (see above), the difference that we observe between the 1980-2005 mean index and the two others cannot be construed to constitute a trend.

SW7608_files/image9.gif

Figure 6. Farnham 1920-2005 mean yearly freeze indexes produced by W5L+ using historical meteorological records. Indexes are divided into four groups that represent the four damaging conditions: extremely low temperature (I1), periods of low temperature (I2), temperature drop (I3), and mild episodes followed by low temperatures (I4).

The separation of the compiled indexes based on the four damaging conditions (fig. 6), shows that conditions I1 (extremely low temperature) and I2 (periods of low temperature) contribute more than I3 (temperature drop) and I4 (mild episodes followed by low temperatures) to the index. In other worlds, conditions I1 and I2 are depicted as generating more damage than conditions I3 and I4 during the period studied. In particular, temperature drop appears to have had little influence on damages occurrences during the past 85 years. This suggests that sudden drops in temperatures were not often a cause of severe damage in Farnham during the period studied. As already mentioned, condition I4 did not generate indexes higher than 2 (see above). This means that only conditions I1 and I2 were recurring causes of major freeze damage in Farnham during the period studied. Neither of these two conditions dominates the other.

W5L+ Application: Future Freeze Risk Projection

A projection of the yearly freeze index for the 1972-2069 period was conducted by running W5L+ with the LARS-WG synthetic meteorological time series. The W5L+ setup remained the same as previously described. The projected indexes for Farnham have been separated into three different periods in figure 7: 1972-2000, 2010-2039, and 2040-2069. Unlike what was observed in figure 6 for indexes generated on past meteorological data, figure 7 shows a continuous and well pronounced negative trend for the index. From a value of 2.5 projected for 1972-2000, the compiled index dropped to 0.8 for 2010-2039 and to less than 0.4 for 2040-2069. This progression suggests that the risk of winter damage to apple trees at Farnham will decline in importance in the future. Extremely low temperature (condition I1) in 2040-2069 will be almost the only potential cause of damage. This means that, despite a rise in average minimum temperatures, extremely low temperatures will continue to occur. This is probably related to the increase in climatic variability projected by the IPCC (2007).

SW7608_files/image10.gif

Figure 7. Farnham 1972-2069 mean yearly freeze indexes produced by W5L+ using LARS-WG projected synthetic meteorological time series. Indexes are divided into four groups that represent the four damaging conditions: extremely low temperature (I1), periods of low temperature (I2), temperature drop (I3), and mild episodes followed by low temperatures (I4).

The W5L+ model predicts that condition I4 (mild episodes followed by low temperatures) will cause no freeze damage in 2040-2069. Rochette et al. (2004) hypothesized that mild episodes followed by low temperatures will be a possible cause of an increase in damaging freezes in the future due to an increase in climatic variability. W5L+ projections do not support this possibility. According to the model, mild temperatures events may increase in length or frequency, but the increase in minimum temperatures will considerably reduce their impacts.

A comparison of compiled indexes of past freeze risk reproduction for 1980-2005 (fig. 6) and those of future freeze risk projection for 1972-2000 (fig. 7) reveals some differences. Despite the fact that they cover very similar periods, the two different W5L+ runs generated indexes of different amplitudes. The difference is mainly due to condition I4. Its contribution to the compiled indexes is much greater when synthetic data are used. This dissimilarity is related to limitations in the meteorological projection and more specifically to limits in the downscaling method. The calculation of condition I4 is based on sequences of daily temperatures (e.g., the number of days in a given period that have a positive average temperature before a fixed low minimum temperature). The problem is that LARS-WG does not specifically take into account the relationships observed between the temperatures of consecutive days for generation of synthetic data. With the I4 results, the downsizing method shows its limits in generating projections that will be sufficiently accurate for the entire set of damaging conditions. This type of deviation can also be observed with the periods of low temperature (I2). The contribution of condition I2 to the compiled index is much lower when synthetic data are used (1972-2000) than when real data are used (1980-2005). Like I4, the calculation of I2 is based on sequences of daily temperatures. These discrepancies do not question the clear trend observed in figure 7, but they underline the need to use W5L+ with different sets of climatic projections that are generated by different downsizing methods.

Conclusions

This study was the first time the original W5L+ model was used to forecast damaging freeze risks during the next 60 years. Results showed that pattern recognition coupled with simple risk index computation has great potential in assessing climate change impacts on freeze occurrence. By differentiating real damaging condition from amplifying factors, W5L+ succeeded in reproducing most of the damaging freeze events that were recorded in the Farnham region between 1921 and 2005. The results generated by W5L+ run on recorded meteorological time series suggested a hierarchy of four pre-identified damaging conditions. During the 1921-2005 period, extremely low temperature (I1) together with periods of low temperature (I2) would have been the main causes of damaging freeze for apple trees in Farnham. Temperature drop (I3) and mild episodes followed by low temperatures (I4) would have had roles in past freeze events, but they would not have been the driving factor for the most important freeze events. As condition I4 generated no risk indexes higher than 2, it is unlikely to be an authentic damaging condition. Used under climatic projections downscaled from CGCM1 simulations using the LARS-WG weather generator, W5L+ results indicate a clear decrease in the damaging freeze risk during the 1972-2069 period. Starting at 2.5 for the 1972-2000 period, the freeze risk index drops to less than 1 for 2010-2039 and to less than 0.5 for 2040-2069. Our findings suggest somewhat different results from those noted previously in the literature. For example, Rochette et al. (2004) predicted increased winter damages in apple trees, while our study found an overall decrease in damages.

Limits associated with the generation of the synthetic meteorological time series make a condition-by-condition analysis of the model projections uncertain. Designing W5L+ required the integration of a snow cover module. The resulting Sn 2 exhibited good performance in reproducing snow cover historical data. It has the advantage of being based on only two commonly measured variables: the maximum daily temperature and the daily amount of snowfall.

The promising performance of W5L+ needs to be confirmed by applying the W5L+ model under a large variety of conditions. Such conditions include different locations and climatic projections issued from different GCM and downscaling techniques. In addition, further development of W5L+ would make the freeze event projection more precise. Decreasing the number of equations and related thresholds on which it is based would be one of the priorities. If simplified, the model would be easier to calibrate and therefore more accurate.

Acknowledgements

We thank the Climate Change Impacts and Adaptation Fund (CCIAF) for providing financial support. We are grateful also to McGill University for assistance in calibrating, validating, and executing the various models.

REFERENCES

Agri-reseau. 2007. Avertissements phytosanitaires / Phytoprotection. Québec City, Québec, Canada: Government of Québec, Agriculture, Pêcheries et Alimentation du Québec . Available at: www.mapaq.gouv.qc.ca. Accessed December 2009.

Audette, M., and S. Ditcham. 2007. Réseau d'essai de cultivars et de porte-greffes de pommiers du Québec - Rapport final du niveau d'introduction, évaluation de cultivars 1997-2003, 1998-2004, et 1999-2005, et essais de porte-greffes 1996-2005, 1997-2005, et 1998-2005. Québec City, Québec, Canada: Centre de référence en agriculture et agroalimentaire du Québec . Available at: www.agrireseau.qc.ca. Accessed March 2010.

Baraer, M., and C. A. Madramootoo. 2007. Can freeze damage impact water demand for crop production in the future? In Proc. USCID 4th Intl. Conf. on Irrigation and Drainage . Denver, Colo.: U.S. Committee on Irrigation and Drainage.

Beck, E. H., R. Heim, and J. Hansen. 2004. Plant resistance to cold stress: Mechanisms and environmental signals triggering freeze hardening and dehardening. J. Biosci. 29(4): 449-459.

Bergeron, L., G. Vigeant, and J. Lacroix. 1997. Chapitre Quebecois de l'etude Pan-Canadienne sur les impacts et adaptation a la variabilite et au changement climatique: Tome V . Québec City, Québec, Canada: Environment Canada.

Burke M. J., L. V. Gusta, H. A. Quamme, C. J. Weiser, and P. H. Li. 1976. Freezing and injury in plants. Ann. Rev. Plant Physiol. 27: 507-528.

Busuioc, A., D. L. Chen, and C. Hellstrom. 2001. Performance of statistical downscaling models in GCM validation and regional climate change estimates: Application for Swedish precipitation. Intl. J. Climatol. 21(5): 557-578.

Charette, R., and R. Krueger. 1992. The low-temperature hazard to the Québec orchard industry. Publication No. 32. Waterloo, Ontario, Canada: University of Waterloo, Department of Geography.

Chouinard C. 1997. Influence du cuivre sur l'endurcissement au froid du porte-greffe de pommier "Malling Merton 106". MSc thesis. Québec City, Québec, Canada: Universite Laval.

Cline, J., R. Gray, and I. Nichols. 2005. Extreme cold temperatures over the past winter likely to cause winter damage to apple trees. Orchard Network 9(2): 9-11.

Coleman, W. K. 1991. Variations in cold resistance among apple cultivars during deacclimatization. J. Exp. Bot. 36(168): 1159-1171.

Eccel, E., R. Roberto, A. Caffarra, and A. Crisci. 2009. Risk of spring frost to apple production under future climate scenarios: The role of phenological acclimation. Intl. J. Biometeorol. 53(3): 273-286.

Hessami, M., P. Gachon, T. B. Ouarda, and A. St-Hilaire. 2008. Automated regression-based statistical downscaling tool. Environ. Modelling and Software 23(6): 813-834.

Hill, H. 1941. Conditions associated with winter injury. Report of the Pomological Society 48: 30-31. Québec City, Québec, Canada: Pomological Society of Québec.

Institut de la statistique Québec. 2007. Production and marketing of apples, by apple growing region, Québec, crop. Québec City, Québec, Canada: Government of Québec. Available at: www.stat.gouv.qc.ca/. Accessed July 2009.

IPCC. 2007. IPCC WGII Fourth Assessment Report: Climate Change 2007: Climate change impacts, adaptation and vulnerability: Summary for policymakers. Geneva, Switzerland: Intergovernmental Panel on Climate Change.

Jonsson, A. M., M. L. Linderson, I. Stjernquist, P. Schlyter, and L. Barring, 2004. Climate change and the effect of temperature backlashes causing freeze damages in Picea abies . Global and Planetary Change 44(1-4): 195-207.

Kaukovirta, E., and P. Syri. 1985. An attempt to reanalyse winter damage of apple trees in Finland during this century. Acta Horticulturae 168: 73-84.

Khanizadeh, S. 2007. Cultural and environmental factors associated with winter injury to apple in northern eastern Canada. Intl. J. Fruit Sci. 7(2): 85-100.

Khanizadeh, S., C. Dubé, Y. Groleau, A. Levasseur, O. Carisse, V. Toussaint, R. Granger and G. Rouselle. 2009. 'SJM167' and 'SJM188', two new winter hardy apple rootstocks. European J. Hort. Sci. 74(5): 223-226.

Kobayashi, K. D. 1983. Modeling cold hardiness of red-osier dogwood. J. American Soc. Hort. Sci . 108(3): 376-381.

Leinonen, I. 1996. A simulation model for the annual freeze hardiness and freeze damage of scots pine. Annals of Botany 78(6): 687-693.

Linden, L. 2001. Re-analysing historical records of winter damage in Finnish apple orchards. Canadian. J. Plant Sci. 81(3): 479-485.

Linden, L. 2002. Measuring cold hardiness in woody plant. Academic diss. Helsinki, Finland: University of Helsinki, Department of Applied Biology.

Palonen, P., and D. Buszard. 1997. Current state of cold hardiness research on fruit crops. Canadian J. Plant Sci . 77(3): 399-420.

Pool, B., T. Martinson, S. Lerch, B. Wilsey, and D. Chicoine. 2004. The Perfect? Freeze. Presented at the Geneva grape field day, Geneva, N.Y.

Quamme, H. A., A. J. Cannon, D. Neilsen, J. M. Caprio, and W. G. Taylor. 2010. The potential impact of climate change on the occurrence of winter freeze events in six fruit crops grown in the Okanagan Valley. Canadian J. Plant Sci. 90(1): 85-93.

QSP. 1870-1953. Annual report of the Pomological Society of Québec. Report published by the Pomological Society of Québec from 1870 to 1953.

Recupom. 2007. Various reports. Québec City, Québec, Canada: Government of Québec. Available at: www.agrireseau.qc.ca. Accessed December 2009.

Rochette, P., G. Belanger, Y. Castonguay, A. Bootsma, and D. Mongrain. 2004. Climate change and winter damage to fruit trees in eastern Canada. Canadian J. Plant Sci . 84(4): 1113-1125.

Russo, J. M., and J. W. Zack. 1997. Downscaling GCM output with a mesoscale model, J. Environ. Mgmt. 49(1): 19-29.

Scott, D., B. Jones, C. Lemieux, G. McBoyle, B. Mills, S. Svenson, and G. Wall. 2002. The vulnerability of winter recreation to climate change in Ontario's Lakelands tourism region. Occasional Paper No. 18. Waterloo, Ontario, Canada: University of Waterloo, Department of Geography.

Semenov, M. A., and E. M. Barrow. 1997. Use of a stochastic weather generator in the development of climate change scenarios, Climate Change 35(4): 397-414.

Semenov, M. A., and E. M. Barrow. 2002. LARS-WG, a stochastic weather generator for use in climate impact studies. User Manual, Version 3.0. Harpenden, U.K.: Rothamsted Research. Available at: www.rothamsted.bbsrc.ac.uk. Accessed March 2010.

Singh, B., and R. Bryant. 2006. Impact et adaptation aux changements climatiques pour les activités de ski et de golf et l'industrie touristique: le cas du Québec. Montreal, Québec, Canada: Ouranos Consortium. Available at: www.ouranos.ca. Accessed March 2010.

Wees, D. 2001. Hardy Fruit Handbook . Publication 45. Montreal, Québec, Canada: McGill University.