High-Resolution Estimation of Summer Surface Air Temperature in the Canadian Arctic Archipelago

15 DECEMBER 2002 ATKINSON AND GAJEWSKI 3601

DAVID E. ATKINSON AND K. GAJEWSKI Laboratory of Paleoclimatology and Climatology, Department of Geography, University of Ottawa, Ottawa, Ontario, Canada

(Manuscript received 27 August 2001, in ®nal form 4 June 2002)

ABSTRACT In the Canadian high Arctic patterns of temperature are poorly resolved at the mesoscale. This issue is addressed using a model to estimate mean summer surface air temperature at high spatial resolution. The effects on temperature of site elevation and coastal proximity were selected for parameterization. The spatial basis is a 1- km resolution digital elevation model of the region. Lapse rates and resultant wind estimates were obtained from upper-air ascents. These were used to estimate the change in temperature with elevation based on the digital elevation model. Advection effects are handled using resultant winds, air temperature above the ocean, and distance to coast. Model results for 14-day runs were compared to observed data. The two effects captured much of the mesoscale variability of the Arctic climate, as shown by veri®cation with point observational data. Sensitivity analyses were performed on the model to determine response to alterations in lapse rate calculation, sea surface temperature, and wind ®eld generation. The model was most sensitive to the lapse rate calculation. The best results were obtained using a moderate lapse rate calculation, moderate wind ®eld, and variable sea surface temperature.

1. Introduction free, which varies signi®cantly on an annual basis. Land surfaces near the coast experience a typical pattern of Interactions of the earth's surface with the atmosphere maritime attenuation, whereas the interiors of larger is- are particularly evident in the Arctic. Plant survival and lands exhibit continental conditions. Exceptions are ar- growth is closely tied to the climate (Arft et al. 1999) eas near snowpacks or extended ice ®elds, which are and the presence of permafrost is a major in¯uence on cooled in the summer. Topographic complexity also con- landscape dynamics (Williams and Smith 1989). Un- tributes to mesoscale variability in temperature, precip- derstanding such climate±surface interactions is impor- itation, and cloudiness. These factors serve to render tant, as future climate changes are predicted to be greater here than in most areas of the world with a potentially questionable results taken from surface air temperature large impact on the landscape (Watson et al. 1995). plots that are based on interpolation from the few avail- However, at the present time, environmental and pa- able meteorological stations. leoenvironmental research in the Arctic is hampered by Improving the spatial resolution of surface air tem- a lack of mesoscale climate, and most importantly temperature estimates is thus an important contribution to perature, data. an understanding of surface climate in this region, and The Canadian Arctic Archipelago (CAA; Fig. 1) is one that is not forthcoming from the existing obser- served by few meteorological stations. The mean sep- vational network. Two mechanisms exist to better un- aration between stations of the Meteorological Service derstand mesoscale temperature: integrating alternate of Canada (MSC) is 500 km and the representativeness data or information into an analysis (Atkinson et al. of all stations suffers due to local coastal bias and highly 2000; Atkinson 2000; Kahl et al. 1992; Alt and Maxwell varied topography and surface types. Physiography con- 1990) or using empirical (Willmott and Matsuura 1995; tributes to temperature pattern variability at the meso- Daly et al. 1994) or physical models (Trenberth 1992) scale for various reasons. The archipelago is heavily of the atmosphere to augment traditional analyses and/ ®orded, exposing land areas to an ocean that can be ice or data sources. covered, contain isolated ¯oating ice ¯oes, or be ice Maxwell (1980, 1982) used information from histor- ical short-term stations and his own experience to sub- jectively modify isotherms to depict cooler ice ®eld/ Corresponding author address: Dr. David E. Atkinson, Geological upland regions. Alt and Maxwell (1990) employed non- Survey of Canada (Atlantic), Bedford Institute of Oceanography, 1 Challenger Dr., P.O. Box 1006, Dartmouth NS B2Y 4A2, Canada. standard, short-term weather observation data from sev- E-mail: [email protected] eral, more recent sources (e.g., Atkinson et al. 2000) to

᭧ 2002 American Meteorological Society

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC 3602 JOURNAL OF CLIMATE VOLUME 15

uation with the underlying surface. The model described below takes as input the synoptic-scale features of the temperature ®eld, as estimated from the MSC upper-air stations, and modi®es their signal using elevational and coastal proximity data derived from the DEM.

2. Data and model description The model was implemented at a spatial resolution of1kmϫ 1 km. Physical processes accommodated in this model areas follows: • The mean environmental lapse rate speci®c to the time period being modeled, derived using temperature data from rawinsonde ascents at MSC upper-air stations in the study region, is used to de®ne the rate of temperature change with elevation. FIG. 1. The CAA. Upper-air stations operated by the MSC are • The mean, low-level wind direction and velocity, de- indicated. rived from rawinsonde ascents, is used to determine the extent to which coastal zones are modi®ed by onshore advective ¯ow. increase spatial detail of a July temperature normal plot • Surface temperatures for locations possessing major for the Queen Elizabeth Islands. Jacobs (1990) linked ice ®elds are stipulated using a linear modi®cation of an automatic weather station to MSC weather stations the base temperature estimate. using transfer functions allowing the generation of data at a ``virtual'' station. Other studies have used the ap- The spatial basis of the model is a DEM of the Ca- proach of guided temperature estimation using a digital nadian Arctic Archipelago, organized as a matrix of 1996 elevation model (DEM) in conjunction with a lapse rate columns by 1833 rows, subset from the U.S. Geological for detailed climate work (Daly et al. 1994; Willmott Survey GTOPO30 DEM of the world (available online and Matsuura 1995; Daly et al. 1997; Daly and Johston at http://edcdaac.usgs.gov/gtopo30/gtopo30.html). Each 1998; Johnson et al. 2000) or to support other types of point represents approximately 1 km2. research (Santibanez et al. 1997; Goodale et al. 1998; The ®rst step in estimating surface air temperature Dodson and Marks 1997). values for each point was to obtain mean environmental In this paper, we describe a semiempirical model of lapse rates for each station. These were generated using the mesoscale summer temperature climate of the Arc- vertical pro®les of dry-bulb temperature obtained from tic. The conceptual basis for the model is that much of twice-daily rawinsonde ascents at stations throughout the spatial variability of the Arctic surface temperature the region (Table 1). The mean ascent curve was de- regime can be accounted for by several processes. Spe- scribed using a ®fth-order polynomial. A high-order ci®cally, we hypothesized that the two most important polynomial was used because it was felt important to contributors to the spatial variability of surface tem- model a shallow, surface inversion that was found to perature patterns at the mesoscale (horizontal scale of be present in many of the ascent pro®les (Figs. 2a,b), tens to hundreds of kilometers) are 1) variation of tem- which are discussed below. perature with elevation, and 2) location with respect to The inversions (Table 1) were smaller in magnitude advective sources of air temperature modi®cation, such than those observed in winter (Bradley et al. 1992; Max- as large bodies of water or ice ®elds. well 1980). Their likely cause is advective, rather than Elevational effects were targeted because many of the radiative, given that the summer net surface radiation islands consist of large central plateaus with a small balance is positive. It was thus assumed that a summer coastal zone. In the northern and eastern parts of the surface inversion at a coastal location is a local-scale archipelago, signi®cant mountainous regions are found. effect that must be removed before using the environ- Concurrent lapse rates applied to site elevations were mental lapse rate to represent interior sites. felt to be the best way to improve estimates of tem- Removal of the inversion involved ®rst detecting the perature in these areas. Advective effects were also in¯ection point on the curve above the inversion using modeled because many of the islands in the CAA are a global-maximum detection algorithm (McCracken and large enough to possess a coast-to-interior heating gra- Dorn 1964). Next, data were extrapolated from this dient that ranges from unimpeded surface heating in the point to the surface using the rate of change that existed interior to coastal locations completely dominated by in the curve above the inversion. The new ascent series maritime air. than had the polynomial equation re®t to it (Fig. 2c). In general, the surface temperature climate at the me- This procedure was veri®ed by comparing estimates of soscale is formed by the interaction of the synoptic sit- surface temperature made by the re®t polynomial to

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC 15 DECEMBER 2002 ATKINSON AND GAJEWSKI 3603

TABLE 1. Upper-air stations used to generate regional estimates of environmental lapse rate. Frequency of inversions observed in mean ascent curves during model runs (1974±88, 1990) are listed. Value class is height of the inversion maximum in m above the ground. Here GT refers to ``greater than 700 m'' (observed only at the Alaska stations). No inver- Ͻ100 Ͻ200 Ͻ300 Ͻ400 Ͻ500 Ͻ600 Ͻ700 GT Upper-air station Lat (N) Lon (W) sion (m) (m) (m) (m) (m) (m) (m) 700 m Alert 82Њ20Ј 62Њ30Ј 10 2 Ð 2 2 Ð Ð 1 Ð Barrow Point (Alaska) 80Њ00Ј 85Њ56Ј 4 Ð Ð Ð 1 5 1 2 4 Barter Island (Alaska) 76Њ14Ј 119Њ20Ј 2 1 Ð Ð 2 2 2 3 5 Cambridge Bay 74Њ43Ј 94Њ59Ј 14 Ð 3 Ð Ð Ð Ð Ð Ð Eureka 63Њ45Ј 68Њ33Ј 4 3 5 5 Ð Ð Ð Ð Ð Iqaluit (Frobisher Bay) 68Њ47Ј 81Њ15Ј 9 1 4 2 Ð 1 Ð Ð Ð Hall Beach 69Њ07Ј 105Њ01Ј 6 3 2 2 2 2 Ð Ð Ð Mould Bay 71Њ18Ј 156Њ47Ј 14 1 1 Ð Ð Ð Ð 1 Ð Resolute Bay 70Њ05Ј 143Њ36Ј 8 2 1 3 1 1 Ð 1 Ð Totals 71 13 16 14 8 11389153 observations from summer research camps at inland niques (e.g., McCullagh 1981; Shepard 1968). Temper- sites (Atkinson 2000). ature values were then obtained by solving the equation Next, the polynomial coef®cients representing the at each grid point using elevation data as the indepen- lapse rate at each station were interpolated throughout dent value. This gave a regionwide estimate of surface the DEM grid. Each coef®cient was interpolated indi- temperature that re¯ected the environmental lapse rate vidually onto the grid using an inverse distance weight without a coastal signal. A concern when using upper- procedure with decay set to a factor of 2; this was se- air temperature data to estimate near-surface air tem- lected to provide a balance between local weighting and peratures is the potential for underestimation of near- range of in¯uence. The paucity of observing sites and surface temperatures; however, consistent bias or large a lack of spatial structure (e.g., no point clustering) did departures from veri®cation data were not observed not warrant use of more specialized interpolation tech- (Figs. 3a±d).

FIG. 2. Typical vertical temperature pro®les for individual ascents showing the inversion (thin line with black dots) and the ®tted polynomial curve (heavier black line): (a) 2 Jul 1987 0000 UTC; (b) 23 Jul 1987 1200 UTC. (c) Mean rawinsonde ascent pro®les for Jul generated by averaging all polynomial estimates for each ascent over the month of Jul for a given year (solid line). Inversion is removed by extrapolating the straight portion of the original curve (dashed line) to the surface (``high-slope'' inversion removal algorithm). All pro®les obtained from the Eureka upper-air station.

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC 3604 JOURNAL OF CLIMATE VOLUME 15

TABLE 2. Wind direction and velocity classi®cation categories. Direction Velocity (Њ true N) Class (km hϪ1) Class 316±45 North 0 Calm 46±135 East 1±13 Low 136±225 South 14±26 Medium 226±315 West 27ϩ High

Temperatures for ice ®elds were then estimated. Ha- vens et al. (1965) demonstrated an average ``ice ®eld cooling'' factor of about 3ЊC using data from two stations, one on top of an ice ®eld and the other nearby on a nonice surface. In the model, ice ®eld locations were assigned a new value that consisted of the initial temperature estimate minus this cooling factor. Next the coastal effect was parameterized. The in¯u- ence of wind for this model is expressed as a mixing of the base land estimates, obtained as described above, with the temperature over the ocean. Wind velocity from the 90-kPa level was extracted from upper-air ascents. The 90-kPa level was selected because it is high enough (ϳ900 m) to be above most topography and to possess the steady characteristics of winds at higher levels, yet low enough to reasonably represent the direction and speed of winds felt at the surface. Based on these values the image was classi®ed into four direction and speed classes, giving a total of 13 categories (12 categories when speeds were Ͼ0, and 1 category for 0 wind speeds) (Table 2). Velocity classi®cation was based on a breakdown of observed wind speeds such that the majority of wind events fell into the ``low'' category and progressively fewer into the ``medium'' and ``high'' categories. These wind categories formed the basis for the selection of a ``matrix ®lter'' (Bonham-Carter 1994) that was applied to a binary representation of the DEM in which land pixels are assigned a value of 1 and ocean pixels are assigned 0. A matrix ®lter is a small, square matrix composed of values that are symmetric and op- posite. This ®lter is placed over a given pixel on the binary DEM. The neighborhood around the pixel that matches the ®lter in size is extracted from the binary DEM and multiplied, pixel by pixel, with the ®lter. All the values in the resulting matrix are then summed to arrive at a single value; this value represents the potential wind in¯uence on a pixel, which is used in Eq. (1) below. The ®lter is arranged such that the largest values are near the middle, representing close proximity to the ocean, with a steady decay to the edge of the ®lter. Thus, pixels near the ocean will feel the greatest potential in¯uence of an onshore ¯ow, decreasing with distance from the coast. The effects of a stronger ¯ow, which are greater potential impact on the near-shore environ- FIG. 3. Daily temperature data observed from an automatic weather station (dashed lines) and estimated by the model for the same lo- ment and farther potential inland penetration, is repre- cation (solid lines) for the years and periods indicated. All plotted sented by a ®lter that has both larger values, to capture data series have been ®ltered using a ®ve-point Gaussian kernel. greater impact, and larger physical size, to represent a Automatic weather station was located at the PCSP Hot Weather greater inland penetration. Greater physical size of the Creek research camp, 30-km inland from the upper-air station at ®lter is used to represent the increased range of effect Eureka on Ellesmere Island.

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC 15 DECEMBER 2002 ATKINSON AND GAJEWSKI 3605

TABLE 3. Model run periods selected for the ``original'' TABLE 4. Sensitivity analyses. parameterization set. Total Year Dates Parameter Nature of alteration runs 1974 9±22 Jul Inversion removal High-slope removal (original) 17 1975 8±21 Jul Low-slope removal 5 1976 4±17 Jul Peak-point removal 5 1977 19 Jul±1 Aug No modi®cation to temperature 5 1978 21 Jul±3 Aug pro®le 1979 9±22 Jul Sea surface temperature Constant over entire region 17 1980 23 Jul±5 Aug (original) 1981 29 Jun±12 Jul Variable over region with low- 5 1982 8±21 Jul slope removal 1983 28 Jun±11 Jul Variable over region with peak- 5 1984 16±29 Jul point removal 1985 6±19 Jul Coastal wind effect Moderate effect application (orig- 17 1986 9±22 Jul inal) 1987 5±18 Jul Maximum effect application with 5 1988 16±29 Jul low-slope removal and 1989 13±26 Jul constant SST 1990 9±22 Jul Maximum effect application with 5 low-slope removal and variable SST of a stronger ¯ow because in the DEM the wind ®lter cannot be applied to an area that it does not physically Three parameterizations were targeted for sensitivity reach. analysis: the inversion removal algorithm, the sea sur- The result of application of the wind ®lter was a face temperature (SST), and the wind effect (Table 4). ``resultant wind effect'' parameter that represented a Each sensitivity combination was tested on 5 separate potential modi®cation of the base temperature estimate years; the same 5 years were used in each case to permit at a site. The maximum value for a resultant wind effect comparison (Table 5). Three additional approaches to is 100, indicating that 100% of the temperature at that dealing with the inversion were considered: a ``peak- pixel is a result of ocean in¯uence, and the minimum point'' removal, in which the slope removal line was is 0, for no modi®cation due to wind. The wind effect drawn vertically down from the point of maximum parameter and the values from the temperature estimates warming to the surface; a ``low-slope'' removal, in image were combined using which the slope removal line was drawn from a point roughly halfway between the original model and the Trsϭ W ϫ T ϩ (1 Ϫ W) ϫ T L, (1) peak-point removal; and ``none,'' in which no alteration to the observed lapse rate was performed. These rep- where T ϭ resultant temperature value at a given point, r resent a gradation in the magnitude of inversion re- W ϭ wind modi®cation value (%), T ϭ air temperature s moval, from a maximum in the original model (``high- over the ocean surface, and T ϭ air temperature over L slope'' removal) to no alteration (none). For SST, the the land surface obtained from the polynomial-based constant value was replaced by values derived from a estimate. Values for T were set at 2ЊC or ranged between s map of mean observed SST (Maxwell 1982). The ex- 0 and 4 C depending on the type of model run being Њ Њ isting wind effect was increased in strength, such that conducted. Topographic modi®cation of wind was not its in¯uence could be felt twice as far inland as in the explicitly parameterized. original model. The ®nal output of a model run was a 1 km 1km ϫ For veri®cation, mean temperature values were cal- grid of estimates of the mean surface air temperature culated using available surface stations present during for the period of the model run. Values were estimated the period of the run. This included both MSC and for all land surfaces over a region encompassing all the islands in the Canadian Arctic Archipelago, Boothia Peninsula, and some of the north coast of the mainland. TABLE 5. Periods for which sensitivity analyses were run. Temperatures were estimated for 2-week averaging periods using an initial set of parameterizations (iden- Period Reason for selecting ti®ed as ``original''), one run for each of 17 years (Table 9±22 Jul 1974 Large zone of negative residual in orig- 3), although the model can be run for any averaging inal model 21 Jul±3 Aug 1978 Lack of inversions for the time period period. Dates of application varied from year to year 28 Jun±11 Jul 1983 Large zone of positive residual in orig- and were chosen to maximize the availability of obser- inal mode vational data (Atkinson 2000) for veri®cation purposes. 16±29 Jul 1984 Large number of stations available for Based on these results various permutations of model veri®cation parameterization changes were run on subsets of 5 years. 16±29 Jul 1988 Climatologically warm summer

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC 3606 JOURNAL OF CLIMATE VOLUME 15

FIG. 4. Model results for 9±22 Jul 1974. Results are estimates of mean surface air temperature (ЊC), integrating the speci®ed time period, arranged ona1kmϫ 1 km grid. Locations of all veri®cation sites have been plotted as black dots. nonstandard data from the Polar Continental Shelf Pro- persistent penetration of cool air southeast from the Arc- ject archives (Atkinson et al. 2000). Model estimates at tic Ocean into the central CAA. grid points coinciding with the station locations were The primary diagnostic tool for assessing model per- extracted and residuals were calculated. Residuals with- formance was a set of residuals obtained by subtracting in the range Ϯ1.4ЊC were considered acceptable, as this model estimates from observed station temperature data. is the minimum standard deviation of two-week means Model estimates were obtained from the grid points obtained from the observational data. Values outside of closest to a given station, and the observed station data this range were mapped to gauge the performance of were averaged for the time period coincident with the the model by revealing regions of systematic over- or model run. These residuals were processed as a com- underestimation. plete set (i.e., and not by year) to obtain a mean absolute error and were both plotted against station distance from coast and elevation to look for situational biases and in 3. Results a mapped form to identify spatial zones of model in- Output from selected years is presented in Figs. 4±6 consistency. An overall mean absolute error of 1.5ЊC and Table 6. Temperature values have been rounded to was obtained on the 386 residual values available for the nearest whole degree Celsius. As expected, cooler the 17 two-week model runs. Considering the residuals temperatures were found at higher elevations in the east- separately for each run, the residuals ranged from having ern Arctic. A general north±south temperature gradient a mean that was close to zero with low variation (e.g., was also captured, as was a ``northwest cool bulge,'' 1976) to a mean that deviated signi®cantly from zero which is a typical temperature pattern caused by the with large variation (e.g., 1974). Overall, negative re-

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC 15 DECEMBER 2002 ATKINSON AND GAJEWSKI 3607

FIG. 5. Model results for 4±17 Jul 1976. Results are estimates of mean surface air temperature (ЊC), integrating the speci®ed time period, arranged ona1kmϫ 1 km grid. Locations of all veri®cation sites have been plotted as black dots. siduals (model overestimation) were more common. All Axel Heiberg Islands (1977, 1978, 1982, 1983, 1984, available residuals were plotted against station elevation 1988, 1990), and often in the eastern parts of Ellesmere and distance from coast (Fig. 7). Biases in the model (1977, 1988, 1990). estimates were not apparent. In general, none of the sensitivity combinations in- A more detailed assessment of model performance vestigated yielded a clearly superior result (Table 8); was determined by considering the size of zones formed however, they all yielded results that were superior to by residuals of a given sign. A large residual zone of the original inversion removal algorithm (high slope). a given sign is more likely a result of a systematic Applying different inversion removal algorithms while shortcoming in the model, whereas small, discontinuous maintaining the constant sea surface temperature re- residual zones of both signs indicate local forcing agents sulted in skewed residual groupings: skewed negative or random ¯uctuations. The 129 residuals that fell out- using the high-slope and low-slope removals, and side the acceptable range formed a total of 57 zones; skewed positive using the peak-point and none resid- those zones, which possessed three or more stations, uals, with the total number of residuals in the acceptable represented only 24% of all observed zones (Table 7). category changing little each time. Similar skewed re- Seven residual zones possessed ®ve or more stations; sults were observed when using the stronger wind ®eld of these six were negative and in all cases the residual with the low-slope removal. The most even distribution zones were situated largely in the northern part of the of residuals was obtained using the low-slope inversion archipelago. Several persistent features were noted in removal with a variable sea surface temperature; how- the residuals plots. Large and small zones of positive ever, it must be noted that this method also yielded one residual were frequent in the north, on Ellesmere and of the lowest numbers of residuals in the acceptable

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC 3608 JOURNAL OF CLIMATE VOLUME 15

FIG. 6. Model results for 19 Jul±1 Aug 1977. Results are estimates of mean surface air temperature (ЊC), integrating the speci®ed time period, arranged ona1kmϫ 1 km grid. Locations of all veri®cation sites have been plotted as black dots.

TABLE 6. Surface air temperature (ЊC) mean and std dev, and sample size of observed values, model estimates, and residuals for each year of ``original'' model run and averaged for all years. Observed Estimate Residual Mean Std dev Mean Std dev Mean Std dev N 1974 5.5 2.6 7.8 1.8 Ϫ2.3 2.5 26 1975 3.4 1.9 4.1 1.3 Ϫ0.6 1.4 27 1976 2.9 1.8 2.9 1.3 0 1.1 23 1977 5.9 2 6 1.6 0 2 27 1978 4.9 1.7 5.4 1.8 Ϫ0.5 2.3 23 1979 4.2 1.8 5 1.5 Ϫ0.8 1.5 26 1980 4.7 1.3 5.3 1.2 Ϫ0.6 1 24 1981 6.4 2.2 7.4 1.6 Ϫ1 2.3 24 1982 6.8 1.6 7.3 1.4 Ϫ0.4 1.8 21 1983 3.5 1.9 2.9 1.8 0.6 1.3 23 1984 5 2.4 4.5 2.1 0.5 1.5 33 1985 5.9 2 7.4 2.2 Ϫ1.5 1.8 24 1986 4.3 2.4 4.8 2.2 Ϫ0.4 1.2 26 1987 6.2 2.6 5.8 2.4 0.4 1.5 14 1988 8.3 2.2 9.6 2.4 Ϫ1.3 2.8 23 1990 6.3 2.1 5.7 2.6 0.6 1.7 22 Overall 5.21 2.4 5.69 2.5 Ϫ0.5 2 386

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC 15 DECEMBER 2002 ATKINSON AND GAJEWSKI 3609

FIG. 7. (a) Residual values for all model runs (n ϭ 386) plotted against distance from coast of the veri®cation station. Distance axis is in km plotted on a log scale. (b) Residual values for all model runs (n ϭ 386) plotted against elevation of the veri®cation station. Elevation axis is in meters plotted on a log scale. range. The largest number of residuals in the acceptable Altering the inversion removal to reduce the mag- range was obtained using no inversion removal; how- nitude of lapse rate correction resulted in the large neg- ever, it also generated the most highly skewed residuals ative residual zones being reduced in size and/or broken set. An examination of the residuals of the sensitivity up into smaller zones (e.g., Figs. 8a±h). Changes intro- analyses shows how the different alterations affected duced by alteration of inversion removal were more speci®c years (Table 9). signi®cant that those resulting from altering the SST or

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC 3610 JOURNAL OF CLIMATE VOLUME 15

TABLE 7. Frequency of occurrence of residual zones possessing certain number of stations, by residual type. Overall Negative residual Positive residual No. of stations No. of zones of No. of zones of No. of zones of in zone this size in plots % this size in plots % this size in plots % 1 34 60 19 56 15 65 2 9 16 3 9 6 27 3 4 7 4 11 Ð Ð 4 3 5 2 6 1 4 5 3 5 2 6 1 4 6 Ð Ð Ð Ð Ð Ð 7 3 5 3 9 Ð Ð 14 1 2 1 3 Ð Ð Total 57 34 23 wind ®elds, which tended to result more in sporadic, signed for the interiors of large islands, thus overesti- low-magnitude changes. mated at these locations. This can be remedied by allowing the model to take into account the size of the land area a given location is situated in, and adjusting 4. Discussion the lapse rate accordingly. Finally, too few wind direc- Areas well represented in the original model included tion options in the wind ®lter was another cause of wind- the central, south-central, and west-southwest regions. related model overestimation at coastal locations. Using When the model was in error it tended to overestimate eight, instead of four, wind directions would improve temperatures. In several years large zones of model this. Another problem that may contribute to a model overestimation were observed in the northwest and overestimation in the ``original'' model runs in the along the eastern edge of the archipelago. Model over- northwest is the value used for the air temperature over estimation occurred in the presence of unusually deep the ocean. That is, 2ЊC is too high for an ocean that is and widespread inversions or when a shallow, surface usually ice covered. This was altered for some of the inversion has undergone a slight surface warming. The sensitivity runs in which a variable air temperature over large zones of systematic overestimation (zones of seven the ocean was used and found to be an improvement. or more stations) accounted for 27% of all residual val- Residuals along the eastern edge of the archipelago ues and were con®ned to four speci®c years in the 17- most likely occurred because none of the upper-air pro- yr run period. When these four years were excluded, ®les are characteristic of this region. Alert, while on the the residuals showed no particular tendency toward pos- coast near the eastern coastal region, is located at the itive or negative skewing. extreme northern limit of this area, which limits the In several cases the model overestimated when the representativeness of its pro®le. Furthermore, the nature resultant wind was zero. This occurred because, without of the interpolation procedure is such that, for much of an onshore wind component, the model did not apply the central-east coast of Ellesmere Island, the in¯uence any cooling to coastal areas. This can be remedied by exerted by Eureka's vertical pro®le, a station poorly allowing some cooling for areas close to the coast even suited to guide estimates in this area, will exceed that during conditions of zero wind. A related problem is of Alert. Incorporation of vertical pro®les from Thule that the model overestimated temperatures on small is- Air Force Base in western Greenland could improve this lands, such as Prince Leopold Island or Seymour Island. situation. In these cases the radiative heating capacity of the small Overestimation occurred more frequently than un- land area of the island is insuf®cient to modify the cool derestimation, and the magnitude of most of the positive lowest levels of the atmosphere. Application of a cor- residuals was small. That fact, coupled with the spatial rected vertical temperature pro®le, which has been de- distribution of residuals, did not indicate systematic

TABLE 8. Residual totals by model factor parameterization set. Factor parameterization set Max wind, Max wind, Variable SST, Variable SST, const SST, variable SST, Residual Original Low slope Peak point None low slope peak slope low slope low slope Less than Ϫ1.4 33 31 18 15 31 16 33 33 Ϫ1.4 to ϩ1.4 76 87 89 92 80 86 83 85 Greater than ϩ1.4 19 19 30 30 26 35 21 19 Total obs 128 137 137 137 137 137 137 137

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC 15 DECEMBER 2002 ATKINSON AND GAJEWSKI 3611

TABLE 9. Surface air temperature (ЊC) mean and std dev, and sample size of observed values, model estimates, and residuals for each year of model run and for all years. Observed Estimate Residual Mean Std dev Mean Std dev Mean Std dev N Low slope 1974 4.9 2.5 6.8 2.1 Ϫ1.9 2.4 23 1978 4.9 1.6 4.9 1.1 0 1.1 25 1983 3.4 1.9 3.1 1.8 0.3 1.2 24 1984 5 2.3 4.2 2 0.9 1.6 34 1988 8.8 2.7 9.6 3 Ϫ0.8 3.1 25 Overall 5.4 2.2 5.7 2 Ϫ0.3 1.9 Peak slope 1974 4.9 2.5 5.4 2.1 Ϫ0.5 1.7 23 1978 4.9 1.6 4.7 1 0.2 1.2 25 1983 3.4 1.9 3 1.7 0.4 1.4 24 1984 5 2.3 4 1.7 1 1.5 34 1988 8.8 2.7 8 2.3 0.8 2.6 25 Overall 5.4 2.2 5 1.8 0.4 1.7 No removal 1974 4.9 2.5 5.3 2.2 Ϫ0.3 1.5 23 1978 4.9 1.6 4.7 1.1 0.2 1.2 25 1983 3.4 1.9 2.9 1.6 0.5 1.5 24 1984 5 2.3 4.1 1.8 0.9 1.4 34 1988 8.8 2.7 7.5 2.4 1.3 2.5 25 Overall 5.4 2.2 4.9 1.8 0.5 1.6 Variable SST with low slope 1974 4.9 2.5 6.6 2.2 Ϫ1.7 2.2 23 1978 4.9 1.6 4.8 1.2 0.1 1.2 25 1983 3.4 1.9 3 1.8 0.4 1.2 24 1984 5 2.3 4.2 2.1 0.9 1.6 34 1988 8.8 2.7 9.7 3 Ϫ0.8 3.1 25 Overall 5.4 2.2 5.6 2.1 Ϫ0.2 1.9 Variable SST with peak point 1974 4.9 2.5 5.3 2.2 Ϫ0.4 1.8 23 1978 4.9 1.6 4.5 1.1 0.4 1.2 25 1983 3.4 1.9 3.4 2.9 0.5 1.3 24 1984 5 2.3 3.9 1.8 1.1 1.5 34 1988 8.8 2.7 7.8 2.3 1 2.5 25 Overall 5.4 2.2 4.9 1.8 0.5 1.7 Max wind with low slope and constant SST 1974 4.9 2.5 6.8 2.1 Ϫ1.9 2.1 23 1978 4.9 1.6 4.8 1.2 0.1 1.1 25 1983 3.4 1.9 3.2 1.8 0.2 1.1 24 1984 5 2.3 4.2 2 0.8 1.6 34 1988 8.8 2.7 9.7 3.1 Ϫ0.9 3.1 25 Overall 5.4 2.2 5.7 2.0 Ϫ0.3 1.8 Max wind with low slope and variable SST 1974 4.9 2.5 6.8 2.2 Ϫ1.8 2.1 23 1978 4.9 1.6 4.8 1.2 0.1 1.1 25 1983 3.4 1.9 3.1 1.9 0.2 1.2 24 1984 5 2.3 4.2 2 0.7 1.6 34 1988 8.8 2.7 9.7 3.1 Ϫ0.9 3.1 25 Overall 5.4 2.2 5.7 2.1 Ϫ0.3 1.8

model underestimation. The most likely situation in there were several instances of underestimation when which positive residuals would occur is during periods reported cloud cover was high. More work is needed to of low cloud cover and low wind speed when surface investigate the relationship between clouds, wind, and heating is greatest. Examination of speci®c situations temperature to ascertain whether the inclusion of cloud revealed that this was not necessarily the case and that cover, perhaps based on wet bulb depression from the

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC 3612 JOURNAL OF CLIMATE VOLUME 15

FIG. 8. Plots of surface air temperature (ЊC) residuals comparing different sensitivity runs for the 1974 run period (9±22 Jul 1974): Sources of the results are as follows (a) original (high- slope algorithm) model; (b) low-slope algorithm; (c) peak-point algorithm; (d) no curve modi- ®cation; (e) low-slope algorithm with variable SST (XVD); (f) peak-point algorithm with variable SST (DVD); (g) low-slope algorithm, constant SST, and maximum wind (XFX); (h) low-slope algorithm, variable SST, and maximum wind (XVX). Shading indicates negative residuals (observed less than model estimates); horizontal line pattern indicates positive residuals (observed greater than model estimates). ascents, should be considered to better estimate in these modi®ed their temperatures (e.g., 1981, 1988). In sev- situations. eral years there were large areas in which the model There were few situations in which the modeled wind apparently overestimated temperature (e.g., 1975, 1976, effect could be assessed due to lack of data for veri®- 1977, 1979), but these large areas consisted of just a cation, although there were instances in which stations few stations, all situated on the coast, that were joined were located within an onshore wind zone that correctly to form a contiguous zone by contouring. These errors

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC 15 DECEMBER 2002 ATKINSON AND GAJEWSKI 3613 are con®ned to coastal regions. This suggests that the temperature data and to explore the physical processes model is capable of generating accurate temperature es- that control the climate in this region. timates for large areas of the Arctic. Another consideration for future work is to expand Sensitivity analyses indicated that modi®cation to the the time periods over which the model can currently inversion removal algorithm had the largest impacts in operate by using vertical temperature pro®les and winds those years for which deep inversions were observed at generated by general circulation models (GCMs) or the multiple stations (1974 and 1988). A successive de- National Centers for Environmental Prediction±Nation- crease in the magnitude of inversion removal resulted al Center for Atmospheric Research (NCEP±NCAR) re- in corresponding decreases in the magnitude and oc- analysis upper-air gridded dataset. A model of this na- currence of negative residuals. This suggests that the ture also provides a means of rendering GCM output in original model algorithm overcompensated when a deep high spatial resolution because the low-resolution, grid- inversion was present, resulting in temperature estimates ded vertical temperature ®elds generated by the GCM that were too large. can be used as surrogate upper-air stations in the model. Although the occurrence of overcompensation was Using gridded pro®les would require detailed work to reduced using inversion removals of lower magnitude, determine how such pro®les should be treated; that is, excessive overcompensation resulted in a problem with coastal or inland. model underestimation. In the ``peak-point'' and The processes selected for parameterization in a mod- ``none'' removal results for 1974, underestimation was el of this nature are dependent on the spatial scale of almost as common as overestimation, but these residuals the model. Use of a higher-resolution DEM would ne- did not form a large, continuous zone similar to that cessitate incorporation of smaller-scale processes, such observed using the original algorithm. In 1988 larger as down-valley winds or slope-differential radiation ef- zones of positive residual appeared, notably in the cen- fects. The current version of this model is designed for tral-north and the west. 1988 was a warm year in the summer use in the high Arctic. For use in other seasons region surrounding MSC Eureka (Edlund and Alt 1989). modi®cations to the handling of inversions would have In general, underestimated sites were situated inland or to be introduced, because the nature of the inversion in sheltered areas, such as at the head of a ®ord. In changes in the wintertime Arctic. Modi®cation to handle 1988, however, sites that were exposed to a stronger the wintertime situation is readily possible, as would be maritime in¯uence, such as that to the east of Ellesmere modi®cation to port the model to other regions of the Island or to the northwest, were still overestimated by world. Work is currently underway for use of the model the model, even in the absence of an inversion correc- in the Yukon. tion, indicating that an erroneous inversion removal al- The model in its present state is limited only by the gorithm was not the only cause of model overestimation. temporal resolution of the input environmental lapse rate A large underestimation that occurred in north Elles- data, in this case, twice-daily upper-air series. It is the- mere Island in 1988 (Biederbick Lake) provides a good oretically possible to generate estimates on a daily or example of the potential problem that can exist with even twice-daily basis; however, estimation at high tem- inland sites. The main estimator for this site is the temporal resolution would possess increased error because perature pro®le from MSC Alert. A location like Alert, of variability in the upper-air series. Estimates that in- unlike a sheltered, inland site, is cooled by two mech- tegrate several days mitigate problems in the upper-air anisms: cool air advection, and blocking of insolation series. Estimation of daily maximum and minimum tem- by low-level cloud. Biederbick Lake, however, expe- peratures could only be performed in conjunction with rienced many days of low wind and cloud, which most some knowledge of the near-surface radiation regime. likely allowed the site to realize its maximum potential It would be possible to get a rough idea of cloud from warming, and which made it different enough from Alert the upper-air data because dewpoint temperature is re- that even a corrected temperature pro®le was unable to corded; however, clouds, and therefore maximum and reproduce its observed temperature. Although small, lo- minimum temperature, have greater spatial variability cal variability will always be a problem in a topograph- than the mean temperature ®eld and the accuracy of ically diverse region such as the High Arctic, the general estimates generated by the model for these parameters mesoscale patterns were captured by the model. would be lower than that for mean temperature. Overall, the model was most sensitive to the manner Future work will improve the model by better param- in which inversions were handled. In terms of area of eterization of the climate of the lower atmosphere and effect, modi®cations to this parameter also had the larg- condition of the surface, and speci®cally the following: est effect. For all combinations of parameterization, the majority of residuals were within the acceptable range. 1) An assessment of the summer inversion regime to Where they exceeded the acceptable range, examination tailor model operation to the different types of in- of their patterns suggested physical causes. The general version that are present; accuracy and physical interpretability of the model re- 2) Improving the near-shore low-wind cooling effect, sults suggest that this is a promising avenue for Arctic so that even under no wind, unless there is a strong climate research, both to generate high spatial resolution wind from the land side, the model should apply

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC 3614 JOURNAL OF CLIMATE VOLUME 15

some sort of coastal cooling to land grid points sit- Their development and use. [Available online at http:// uated beside ocean points; www.ocs.edu/prism/prisguid.pdf.] ÐÐ, R. P.Neilson, and D. L. Phillips, 1994: A statistical±topographic 3) Improving the spatial distribution of air temperature model for mapping climatological precipitation over mountain- estimates above the ocean by incorporating a more ous terrain. J. Appl. Meteor., 33, 140±158. detailed sea surface temperature or mean sea ice con- ÐÐ, G. H. Taylor, and W. P. Gibson, 1997: The PRISM approach ditions map; to mapping precipitation and temperature. Preprints, 10th Conf. on Applied Climatology, Reno, NV, Amer. Meteor. Soc., 10±12. 4) Using the vertical pro®le of dewpoint temperature Dodson, R., and D. Marks, 1997: Daily air temperature interpolated in conjunction with the dry-bulb temperature pro®le at high spatial resolution over a large mountainous region. Cli- to get an estimate of clouds; and mate Res., 8, 1±20. 5) Incorporating a more detailed ice ®eld/glacier map Edlund, S. A., and B. T. Alt, 1989: Regional congruence of vegetation and model downslope katabatic effects in large gla- and summer climate patterns in the Queen Elizabeth Islands, Northwest Territories, Canada. Arctic, 42, 3±23. cial valleys. Goodale, C. L., J. D. Aber, and S. J. Ollinger, 1998: Mapping monthly precipitation, temperature, and solar radiation for Ireland with In general the model has performed satisfactorily and polynomial regression and a digital elevation model. Climate physical mechanisms can be suggested for most errors. Res., 10, 35±49. This study has shown that a signi®cant amount of the Havens, J. M., F. MuÈller, and G. C. Wilmot, 1965: Comparative me- mesoscale climatic variability of the Arctic may be ex- teorological survey and a short term heat balance study of the plained by a few factors operating on the complex to- White Glacier, summer 1962. McGill University, 68 pp. Jacobs, J. D., 1990: Integration of automated station data into ob- pography of the Arctic. jective mapping of temperatures for an Arctic region. Climatol. Bull., 24, 84±96. Acknowledgments. This paper is a contribution to the Johnson, G. L., C. Daly, G. H. Taylor, and C. L. Hanson, 2000: Spatial Climate Systems History and Dynamics (CSHD) project variability and interpolation of stochastic weather simulation model parameters. J. Appl. Meteor., 39, 778±796. funded by the Meteorological Service of Canada (MSC) Kahl, J. D., M. C. Serreze, and R. C. Schnell, 1992: Tropospheric and the Natural Science and Engineering Council of low-level temperature inversions in the Canadian Arctic. Atmos.± Canada (NSERC). The DEM data are distributed by the Ocean, 30, 511±529. EROS Data Center Distributed Active Archive Center Maxwell, B., 1980: The Climate of the Canadian Arctic Islands and (EDC DAAC), located at the U.S. Geological Survey's Adjacent Waters: Climatological Studies No. 30. Vol. 1. At- mospheric Environment Service, Environment Canada, 532 pp. EROS Data Center in Sioux Falls, South Dakota. ÐÐ, 1982: The Climate of the Canadian Arctic Islands and Adjacent Waters: Climatological Studies No. 30. Vol. 2. Atmospheric En- vironment Service, Environment Canada, 589 pp. REFERENCES McCracken, D., and W. Dorn, 1964: Numerical Methods and Fortran Programming. John Wiley and Sons, 457 pp. Alt, B. T., and J. B. Maxwell, 1990: The Queen Elizabeth Islands: McCullagh, M. J., 1981: Creation of smooth contours over irregularly A case study for Arctic climate data availability and regional distributed data using local surface patches. Geogr. Anal., 13, climate analysis. Canada's Missing Dimension: Science and His- 51±63. tory in the Canadian Arctic Islands, C. R. Harrington, Ed., Can- Santibanez, F., L. Morales, J. Delafuente, P. Cellier, and A. Huete, ada Museum of Nature, 294±326. 1997: Topoclimatic modeling for minimum temperature predic- Arft, A. M., and Coauthors, 1999: Responses of tundra plants to tion at a regional scale in the central valley of Chile. Agronomie, experimental warming: Meta-analysis of the International Tundra 17, 307±314. Experiment. Ecol. Monogr., 69, 491±511. Shepard, D., 1968: A two-dimensional interpolation function for ir- Atkinson, D. E., 2000: Modeling July mean temperatures on the Fosh- regularly-spaced data. Proc. 1968 Association for Computing eim Peninsula, Ellesmere Island, Nunavut. Environmental Re- Machinery (ACM) National Conf., Princeton, NJ, ACM, 517± sponse to Climate Change in the Canadian High Arctic, Geo- 524. logical Survey of Canada Bull., No. 529, Geological Survey of Trenberth, K. E., Ed., 1992: Climate System Modeling. Cambridge Canada, 99±111. University Press, 788 pp. ÐÐ, B. T. Alt, and K. Gajewski, 2000: A new database of high Watson, R. T., M. C. Zinyowera, and R. H. Moss, Eds., 1995: Climate arctic climate data from the Polar Continental Shelf Project ar- Change 1995: Impacts, Adaptations and Mitigation of Climate chives. Bull. Amer. Meteor. Soc., 81, 2621±2629. Change: Scienti®c±Technical Analyses. Cambridge University Bonham-Carter, G. F., 1994: Geographic Information Systems for Press, 879 pp. Geoscientists: Modelling with GIS. Pergamon Press, 398 pp. Williams, P. J., and M. W. Smith, 1989: The Frozen EarthÐFun- Bradley, R. S., F. T. Keimig, and H. F. Diaz, 1992: Climatology of damentals of Geocryology. Cambridge University Press, 306 pp. surface-based inversions in the North American Arctic. J. Geo- Willmott, C. J., and K. Matsuura, 1995: Smart interpolation of an- phys. Res., 97 (D14), 15 699±15 712. nually averaged air temperature in the United States. J. Appl. Daly, C., and G. L. Johnson, cited 1998: PRISM spatial climate layers: Meteor., 34, 2577±2586.

Unauthenticated | Downloaded 09/28/21 05:03 AM UTC