Building an indicator to characterize the thermal conditions for plant growth on an Arctic archipelago, Svalbard

Ecological Indicators 66 (2016) 623–631

Contents lists available at ScienceDirect

Ecological Indicators journal homepage: www.elsevier.com/locate/ecolind

Building an indicator to characterize the thermal conditions for plant growth on an Arctic archipelago, Svalbard Daniel Joly a,∗ , Geir Arnesen b , Eirik Malnes c , Lennart Nilsen d a

Université de Franche-Comté, CNRS, Laboratoire THEMA, 32, rue Mégevand, Besanc¸on, France Ecofact Nord AS, Grønnegata 65, N-9008 Tromsø, Norway Northern Research Institute, P.O. Box 6434, Tromsø Science Park, N-9294 Tromsø, Norway d University of Tromsø, The Arctic University of Norway, Department of Arctic and Marine Biology, Dramsveien 201, N-9037 Tromsø, Norway b c

a r t i c l e

i n f o

Article history: Received 31 July 2015 Received in revised form 13 November 2015 Accepted 1 December 2015 Available online 3 March 2016 Keywords: Temperature Interpolation Modelling Arctic Svalbard

a b s t r a c t Plant growth in the Arctic is strictly dependant on thermal conditions. The purpose of our study is therefore to calculating temperature distributions on the Svalbard archipelago at a relatively high spatial resolution. The model is designed to reflect both the length of the growing season and the temperature sum for a given area (i.e. growing degree-days (GDD)). GDD on Svalbard is defined as the cumulative sum of positive mean daily temperatures in the months of June, July and August. The temperature distribution of GDD for the entire archipelago is calculated from both local and regional information. Local information is derived from data collected in a small area in northwestern Spitsbergen (Kongsfjorden) where a network of 45 thermal sensors recorded air temperatures for five years (2001–2005). A local GDD parameter is computed by a linear combination of elevation, valley depth and NDVI (normalized difference vegetation index). Then this local GDD is applied to the whole of Svalbard (GDD1) and refined stepwise by adding environmental variables such as cloud fraction, land surface temperature, sea surface temperature, distance to the ocean and number of snow-free days. Because the official network of climatological stations on Svalbard is not dense enough and sufficiently well-distributed across the archipelago to enable spatial interpolations for four years only (2011–2014), all outputs are statistically evaluated and adjusted using the values recorded at 9 (2011), 12 (2012) and 13 (2013–2014) meteorological stations (GDDref ) and used as a set of evaluation data. The final model (GDDmean ), which is the mean of the annual models estimated by regression (GDDest ), performs well: the central parts of Spitsbergen, known for its comparatively high temperatures, contrast with the colder northern and eastern parts of the archipelago. © 2015 Elsevier Ltd. All rights reserved.

1. Introduction The Arctic vegetation experience changing environmental conditions including higher temperatures induced by global change (Bliss, 1971; Callaghan et al., 1999; Lenoir et al., 2008). It is likely that we will see changes in the vegetation cover in near future, as species adapted to high arctic conditions may suffer decreasing populations, while others may be able to colonize habitats which has been too hostile until now. These spatial redistributions of plants patterns depend on many geographical features such as soil composition, texture and moisture, topography and climate elements, etc. Even if temperature is one of the most limiting factors

∗ Corresponding author at: UiT, Université de Franche-Comté, CNRS, Laboratoire THEMA, 32, rue Mégevand, Besanc¸on, France. Tel.: +33 381665402. E-mail addresses: [email protected] (D. Joly), [email protected] (E. Malnes), [email protected] (L. Nilsen). http://dx.doi.org/10.1016/j.ecolind.2015.12.005 1470-160X/© 2015 Elsevier Ltd. All rights reserved.

of plants in the Arctic, energy alone does not help if, according the “Liebig’s Law of Minimumm” first described by Sprengel in 1828, other conditions are below any minimum thresholds (water, nutrients, substrate, etc.). The problem is that, except for temperature, we have no other data available at large scale. That is why this work is only based on temperature data that enables us to have knowledge on how temperature is structured locally and to understand the thermal conditions in which plants currently grow. The large-scale predictive temperature models developed by IPCC (IPCC, 2013) are by far too coarse to be useful in this context. Looking back to the last decades, mean temperatures during the summer months have often been used to define bioclimatic zones in both boreal and arctic regions characterized by distribution of vegetation units (Elvebakk, 1994; Moen, 1998; CAVM team, 2003; Karlsen et al., 2005; Walker et al., 2005) or single species of indicative value (Thuiller et al., 2004). However, in Svalbard, temperatures in most areas are poorly documented, and especially so in the remote eastern, northern and southern reaches of the

624

D. Joly et al. / Ecological Indicators 66 (2016) 623–631

archipelago. Hence, the correct distribution of the bioclimatic zones in the archipelago has been a matter of discussion for years due to mostly rather subjective approach, and the lack of empirical temperature data. This state of affairs detracts from our understanding of Svalbard’s and the links between temperature and plants are not explained or validated. Over the last years, several new automatic weather stations in Svalbard is yielding temperature data, improving the access to empirical temperature series. However, Svalbard will probably never be covered with a dense network of weather stations, and it will not be possible to map distribution of temperature on a local scale from such data alone. When empirical data is not available or the quality is not satisfactory, a natural approach is to use a model to fill in the gaps. The objective of our work is thus to add some objectivity and accuracy to the knowledge of temperature distribution in Svalbard by presenting a geographical temperature model of the archipelago. We have chosen to use the parameter “growing degree days” (GDD), here defined as the sum of positive mean daily temperatures in the months of June, July and August. Some studies also refer to the same parameter as effective temperature sum’ (Tuhkanen, 1993; Karlsen, 2003) which has been proved slightly better correlated to distribution of plants and vegetation than for example the mean July temperature, which also has been widely used. Our starting point was modelling distribution of GDD in the Kongsfjorden area of western Svalbard using available data from a network of 45 temperature sensors distributed throughout the area (Joly et al., 2010; Nilsen et al., 2013b). The objective is here wider and consists to develop a model for calculating temperature distributions throughout the Svalbard archipelago at high spatial resolutions similar to what was achieved in Kongsfjorden. The model should then be able to reflect both the variations in the length of the growing season as well as other more large-scale parameters affecting the GDD. The method for computing GDD for the entire archipelago was based on a series of regressions, and the first explanatory variable was the extrapolation of the local model developed in Kongsfjorden to every 100 × 100 m pixel of the archipelago. Four other explanatory variables are retained to take into account the temperature variation on a regional scale. The ‘ground surface heating potential index’ (GSHPI) reflects the thermal conditions under which plants actually grow. ‘Sea surface temperature’ (SST) and ‘distance to the open

sea’ (DOS) are intended to capture the thermal influence of the sea on air temperature, which is highly dependent on the North-East Atlantic current and the relatively continental character of central Svalbard. The ‘number of snow-free days’ (NSFD) provides guidance as to the potential length of the growing season that starts as soon as the snow melts. The output was evaluated by comparing estimated results with empirical data from weather stations, and adjustments were done in order to improve the performance. The final result is presented as a map of Svalbard depicting the spatial variation of GDD at a spatial resolution of 100 m.

2. Study area and material 2.1. Svalbard The target area to be modelled was the entire archipelago of Svalbard except for the island of Bjørnøya, which although part of the archipelago, lies far to the south. Svalbard’s three main islands (Spitsbergen, Nordaustlandet and Edgeøya) lie between 76.5 and 80.5◦ north, about 1000 km from the North Pole and astride the boundary between the Barents Sea and the Arctic Ocean (Fig. 1). The total land area of the archipelago is approximately 63,000 km2 with its largest island, Spitsbergen, covering some 38,000 km2 . Except a low coastal line called “Strandflat” which elevation do not exceed 100 m, most of Svalbard consists in mountains (maximum altitude 1717 m at the Newtontoppen) with steep slopes and large glaciers (Table 1). Meteorological data show that Svalbard has a polar oceanic climate with cool and wet summers. At low elevations, mean summer temperatures vary, depending on the year, from −1 ◦ C, 1 ◦ C in the northern and eastern areas (Karl XII Øya, Kvitøya) to 5 ◦ C, 7 ◦ C in the central parts of Svalbard (Table 2). A positive thermal gradient occurs between the west coast (Hornsund, Sørkapp, Ny-Ålesund) and the central part of Spitsbergen (Sveagruva, Longyearbyen, Pyramiden) with mean summer temperatures of 4.5 ◦ C and 6.1 ◦ C respectively in 2013–2014 (http://eklima.met.no). The Meteorological Institute of Norway operated meteorological stations for an extended time at only a very few locations before 2010 (Table 1). Accordingly, the study concentrates on the years 2011 to 2014, which are the only ones in which enough stations

Fig. 1. The location of the Svalbard archipelago in the Arctic and locations of the weather stations. The numbers refer to those of Table 2. The square with number 7 on the north-west coast of Spitsbergen marks the Kongsfjorden area where the 45 temperature loggers were installed.

D. Joly et al. / Ecological Indicators 66 (2016) 623–631

625

Table 1 Frequency (%) of pixels for nine elevation (m) ranges.

%

<50

50–99

100–199

200–299

300–399

400–499

500–749

750–999

>=1000

12

7

12

13

14

12

21

6

3

Source: DTM at 100 m resolution built by the Norwegian Polar Institute.

Table 2 Mean summer temperatures for the years 2007–2014. Data from the 13 stations were made available by the Norwegian Meteorological Institute (http://eklima.met.no). Stations

Elev.

1-Karl XII Øya 2-Kvitøya 3-Verlegenhuken 4-Ny-Ålesund 5-Kongsøya 6-Pyramiden 7-Longyearbyen (airport) 8-Edgeøya 9-Sveagruva 10-Akseløya 11-Hornsund 12-Hopen 13-Sørkapp

5 10 8 8 20 20 28 14 9 6 10 6 10

2007

2008

2009

2010

2011

.6

5.4

3.9

4.2

4.2

6.7 1.7 4.6

5.2 1.7 4.8

5.9

5.2

5.3

4.8

3.9 2.7

3.8 1.9

4.3 2.3

2.5

operated for statistics to be done. Most of the stations are located on the west coast or inner fjord zone of Spitsbergen (Fig. 1). Additionally a handful of automatic stations are located in the eastern and northern parts of the archipelago. GDDref is calculated by summing the positive daily mean temperatures recorded by the 9 (2011), 12 (2012) and 13 (2013–2014) meteorological stations. 2.2. The Kongsfjorden area The local GDD model was constructed from data collected from a small area (approximately 300 km2 ) around the inlet of Kongsfjorden in northwestern Spitsbergen (Fig. 1). There is an obvious climatic gradient from the wind-exposed plain in the western part of the peninsula (Kvadehuksletta) with its polar desert-like appearance to the more sheltered areas in the east with their more extensive vegetation cover. The meteorological station of NyÅlesund is situated half way along this oceanic–inland gradient and has a mean July temperature of 4.9 ◦ C and mean annual precipitation of 385 mm (Førland et al., 1997, 2011). Our network of 45 air temperature loggers (Joly et al., 2010) was laid out so as to cover relevant gradients created by elevation, aspect, slope, proximity to coastline and proximity to glaciers. Since the aim was to study temperatures in relation to the distribution of vegetation, the 45 sensors were positioned 10 cm above ground level and protected from direct sunlight and precipitation by 20 cm × 20 cm shields. The 64 kB memory of the Hobo H8Pro temperature loggers with external temperature sensors enabled them to store up to 32,520 readings. They were programmed to make two readings per hour and recorded data from June 2001 to August 2005.

2.5 5.5

6.4 1.7 5.2 4.3 3.9 2 2.2

2012

2013

2014

.5 0.8 3.2 4.9 1.5

1.1 1.7 4.6 4.9 2.2 6.5 6.6 4.3 6.3 5.7 4.9 4.5 4.7

−1 −0.5 1.3 4.1 0.7 5.7 6 3.2 5.6 5.3 4.4 3.2 4.3

5.6 2.3 4.8 4.5 3.6 2.6 3.1

DTM at 100 m resolution built by the Norwegian Polar Institute: elevation, slope gradient, aspect broken down into sine and cosine, rugosity, exposure index,1 solar radiation, distance to the nearest ridge and river line, distance to the open sea and to the fjord, index of thermophily (Elvebakk, 1990). NDVI is the twelfth explanatory variable. Each of them was determined by GIS for each of the logger locations (Joly et al., 2010). All the significant explanatory variables at the 5% significance level (p < 0.05) were included in the multiple regression function. Processing of the 45 GDD1ref values revealed that elevation explains 86% of GDD1ref variance (r = −0.93) (Nilsen et al., 2013b). Four other variables also explain spatial variations in Gdd1ref although to a lesser extent: Index of thermophily (0.45), distance to nearest river thalweg (0.43), NDVI (−0.42) and exposure index (r = −0.4). In the context of the current study, the aim of which is to model GDD throughout the Svalbard archipelago, it would have been unreasonable to use the same five explanatory variables used in the local study, hence, some adjustments were done before extrapolating to the entire archipelago. First, only the variables available for the whole Svalbard are retained. Second, the variable ‘Distance to the nearest river thalweg’ was replaced by the more integrative variable ‘valley depth’, which better takes into account the protection of the topographical depressions against winds (Joly et al., 2012). Third, because the colinearity between elevation and exposure is much higher for the whole of Svalbard than for Kongsfjorden alone, exposure has not been retained in this study. Then, by using multiple stepwise regression, the GDD1 model for the whole of Svalbard was estimated by the following function: GDD1 = 200 + (−0.29 Elev) + (0.18 Dval) + (21.1 NDVI)

(1)

3. Method 3.1. Local GDD model The average daily temperatures of 460 days (92 days per summer × 5 summers from 2001 to 2005) were calculated for each of the 45 stations in the Kongsfjorden area. The 45 GDD values (GDD1ref ) were then calculated by adding the positive mean daily temperatures for the months June, July and August. The GDD values for the entire Kongsfjorden area (GDD1est ) were estimated by a multiple linear regression with GDD1ref as the response variable. Eleven explanatory variables are extracted or calculated from the

where Elev is elevation in metres, Dval is valley depth and NDVI is Normal Difference Vegetation Index graded into nine homogeneous classes. The regression coefficient indicates that GDD1 falls by 0.29 ◦ C for each vertical increment of 100 m in elevation; lower Dval and higher NDVI yield a higher GDD.

1 A deeply incised valley or a narrow valley bottom takes a negative value whereas a high point (summit of a hill, crest line) has a positive value.

626

D. Joly et al. / Ecological Indicators 66 (2016) 623–631

Table 3 Summary of explanatory variables (GDD1, GSHPI = ground surface heating potential index, SST = sea surface temperature (◦ C), DOS = distance to open sea (km), and NSFD = number of snow-free days). GDD values (predicted variable) for the locations of the meteorological stations and GDD for the four years under consideration are indicated. Stations

GDD1

GSHPI

SST

DOS

NSFD

GDD 2011

GDD 2012

GDD 2013

GDD 2014

Hornsund Sveagruva Sørkapp Ny-Ålesund Hopen Longearbyen Verlegenhuken Edgeøya Karl XII Island Akseløya Kongsøya Kvitøya Pyramiden

273 290 309 268 281 268 268 289 255 260 241 222 294

22.4 29.0 17.5 32.8 18.4 33.6 37.6 19.2 5.7 16.7 14.7 4.2 40.0

23.7 25.0 16.9 70.6 0 64.7 0 0 0 58.2 0 0 51.0

3.1 37.6 2 15 0.1 47.8 2 0.2 0.2 2.44 5 0.2 60.9

83 72 30 67 30 101 40 47 30 81 30 30 89

345 466 193 486 191 573 256 154

312 423 280 427 247 492 274 211 51 388 172 120

429 561 429 435 402 585 327 376 115 505 192 185 598

405 515 399 375 297 552 166 310 32 492 116 42 524

3.2. Large scale environmental factors The GDD1 contains variables mainly significant on a local scale (Joly and Brossard, 2007). In order to address large-scale variation affecting temperature new variables had to be introduced. 3.2.1. Ground surface heating potential index (GSHPI) Temperature on the ground surface, different from that of air measured at 2 m above the ground, is controlled by many factors of the soils and its thermal conductivities, i.e. heat storage potentials which in turn are related to grain and pore sizes, organic matter content, etc. In fact, temperature on the ground surface is an indicator which synthesizes all these numerous factors. It is a good proxy of the potential given by the environment to allow the plants to grow optimally. GSHPI was modelled using data from the MODerate resolution Imaging Spectroradiometer (MODIS) sensors on-board the Terra and Aqua satellites. MODIS images are the products with the most suitable time resolution because the near polar orbit provides two scans of Svalbard each day. However, the high cloud fraction over Svalbard makes it impossible to establish pixel-by-pixel time series of the available land surface temperature (LST) products. Hence, a maximum temperature image (LST-MAX) was made, using overlay procedures retaining the highest value of each image product. The resulting image indicated the temperature of each pixel area on the warmest cloudless day in the time span 2002–2008. It provides a reasonable estimate of the ground surface warming potential. The warmest places are located in central Spitsbergen (values >8 ◦ C). However, LST-MAX was not expected to be closely correlated with GDD, as warming of the ground surface is only efficient during sunny weather. In order to be useful, the LST-MAX variable had to be adjusted using cloud fraction products. The assumption is that, in the absence of clouds, solar radiation heats the material at the earth’s surface, which in turn warms the layer of air close to the ground surface. Conversely, where the ground surface does not receive direct radiation from the sun, no heating occurs and the air remains homogeneously cold or cool in horizontal and vertical directions. Thus, higher ground temperatures should logically develop in areas where cloud cover is less frequent compared with locations with frequent cloud cover. The monthly mean cloud fraction data for June, July and August were used: 2002–2008 (MODIS Terra – 20 files) and 2005–2008 (MODIS Aqua – 12 files). Each monthly product is a mean of the daily products for the specific month. The mean of the 32 available monthly products was calculated to obtain an overall cloud cover trend image. GSHPI was calculated from the equation: GSHPI = LST × (1 − cloud-fr)

(2)

369

3.2.2. Sea surface temperature (SST) Data on sea surface temperature (SST) with 4 km resolution were acquired from the advanced very high resolution radiometer (AVHRR) sensor on board the NOAA polar-orbiting satellites. The imaging data were reclassified and transformed using a fifth degree polynomial adjustment (Bolstad, 2012). The resulting trend surface image was applied to the land pixels. 3.2.3. Distance to the open sea (DOS) In the local study carried out on the west coast of Spitsbergen, a west-to-east temperature gradient was identified. The mean temperature rises with distance from the ocean under the influence of continentality, which strengthens towards the centre of the archipelago (Joly, 1994; Joly et al., 2010). The same pattern can be observed for the major fjord systems along the west coast of Spitsbergen (Ørbæk et al., 1999). It is likely that a similar gradient can be found, from east-to-west, along the east coast. For every pixel, the value of DOS (km) is the Euclidean distance to the nearest ocean or sea. 3.2.4. Number of snow-free days (NSFD) The length of the growing season is closely correlated with the time from snowmelt in spring until the snow covers the land again in autumn. The number of snow-free days expresses the effect of the precipitation and temperature factors, which are difficult to assess empirically at a fine-scale resolution. Data on snow fraction were obtained from the MODIS EOS Terra sensor. The level 3 global 500 m grid daily mosaic snow product (MOD10A1 version 5) contains the snow cover fraction (SCF) product. Data were available for 10 summer seasons (5 March 2000 to 1 October 2009). The first step was to produce daily snow cover maps. In order to fill in pixels that were cloud covered for more or less long periods, linear interpolations from cloud free days were performed. According to Nilsen et al. (2013a), the first day of snow-free conditions was defined as the first day with SCF < 10% followed by six consecutive days with SCF < 10% after April 1. The first day of snow cover in autumn was defined as the first day with SCF > 50% followed by six consecutive days with SCF > 50% after September 1. The snow period was calculated as the average count of snow days over the years for which data was available. 3.3. Regression process and quality of variables Table 3 summarizes the values of the variables introduced in the regression process. The response variables (GDDref 2011–2014) vary in large proportions. On average, the coldest summers were 2012 (GDDref = 328) and 2014 (368). The summer of 2013 was

D. Joly et al. / Ecological Indicators 66 (2016) 623–631 Table 4 Matrix of correlation between explanatory variables (summers 2013 and 2014).

GDD1 GSHPI SST DOS NSFD

GDD1

GSHPI

SST

DOS

NSFD

1

0.59 1

0.35 0.55 1

0.39 0.68 0.61 1

0.42 0.62 0.83 0.72 1

the warmest (GDDref = 451). As regards the explanatory variables, GDD1 exhibits low spatial variability (minimum = 222 at Kvitøya and maximum = 312 at Svalbard Airport). That is because the 13 meteorological stations are all located at low elevations (Table 1). GDDref clearly highlights the thermal difference between the stations bordered by cold sea and those located on the west coast, which benefit from the mild North Atlantic Drift. The distance to the ocean mainly contrasts stations located in the heart of Spitsbergen (Svalbard Airport, Pyramiden and to a lesser extent Sveagruva) with stations close to the open sea. Finally, NSFD contrasts stations free of snow for a long time (Svalbard Airport, Pyramiden) with cold northern (Karl XII Øya), southern (Sørkapp) and eastern (Kvitøya, Kongsøya) stations, which are snow-free for just 30 days on average. 3.3.1. Multi-colinearity The matrix of correlation between explanatory variables (Table 4) shows that multiple colinearities arise. For example, SST shares 68% of variance in common with NSFD (r = 0.83). In this case, it may be that NSFD adds little to the estimation once SST has been included in the regression. To ascertain this, two colinearity tests were run. The first test was for tolerance, which is the share of variance of one variable that is not explained by the other variables. A high tolerance value corresponds to weak colinearity. Table 5 reveals that the 0.2 threshold taken to be the limit below which colinearity is critical (Hair et al., 2006) was rare and only DOS was below this value in 2011. The second test was for the variance inflation factor (VIF), indicating the rise in the variance of coefficients in the presence of colinearity in a least squares regression. Values of 10 (Hair et al., 2006), 5 (Rogerson, 2001) or 4 (Pan and Jackson, 2008) have been recommended as the maximum levels of VIF. All the values of VIF are below 5, except for DOS in 2011 (Table 4). VIF confirms the tolerance values. Consequently, we removed DOS from the 2011 regression. 3.3.2. Significance of the explanatory variables Table 6 shows that most of the explanatory variables are significant at the 1% level (SST, DOS, and NSFD). Conversely, GSHPI (r = 0.57, p = 0.107) is less significant in 2011, only nearly reaching a 10% level, but highly significant at the 1% level for 2012 and 2013. GDD1, with r from 0.07 (2011; p-value = 0.86) to 0.68 (2014; Table 5 Tolerance and VIF: two indicators for assessing multi-colinearity. GDD1

GHSPI

SST

DOS

NSFD

Tolerance 2011 2012 2013–2014

0.38 0.65 0.64

0.41 0.49 0.39

0.36 0.34 0.31

0.16 0.48 0.39

0.30 0.27 0.23

VIF 2011 2012 2013–2014

2.65 1.54 1.56

2.47 2.03 2.55

2.76 2.96 3.23

6.15 2.07 2.53

3.38 3.76 4.29

Bold figures indicate a high co linearity.

627

Table 6 Significance test of the explanatory variables. GDD1

GHSPI

SST

DOS

NSFD

p-Value 2011 2012 2013 2014

0.864 0.067 0.018 0.011

0.107 0.003 0.005 0.021

0.003 0.008 0.005 0.003

<0.001 0.008 0.016

0.004 <0.001 <0.001 <0.001

r 2011 2012 2013 2014

0.07 0.54 0.64** 0.68**

0.57 0.78*** 0.73*** 0.63**

0.84*** 0.84*** 0.72*** 0.75***

** ***

0.72*** 0.70*** 0.65**

0.85*** 0.83*** 0.81*** 0.82***

Significant at the 5% level. Significant at the 1% level.

p-value = 0.01) is highly unstable. The p-value and the r of GDD1 for 2011 mean the variable cannot be retained because (i) there is a high probability that the deviating results are uncertain, and (ii) there is no statistical connection between GDD1 and GDD2011 ref . However, GDD1 is retained for the years 2012–2014, as it is so close to the 5% significance threshold in 2012 and below the threshold for 2013 and 2014. Given the different indicators, we considered that all explanatory variables could be retained for the multiple regressions (except for GDD1 and DOS for 2011), although it could be debated whether there was any point in retaining certain weak variables (GDD1 because of its low r; SST and NSFD because of the colinearity between them). 3.3.3. Regressions The GDDref values monitored at 9–13 stations throughout Svalbard (response variable) were estimated by five (three in 2011) explanatory variables using a linear regression method (XLSTAT). The GDDref value at a pixel i,j was, then, predicted from the following form: GDDref[i,j] = intercept + (GDD1i,j × CRGDD1 ) + (GHSPIi,j × CRGHSPI ) + (SSTi,j × CRSST ) + (DOSi,j × CRDOS ) + (NSFDi,j × CRNSFD )

(3)

where CR is the regression coefficient; GDD1i,j is GDD1 at pixel i,j; GHSPI is the ground surface heating potential index built using cloud cover and maximum land surface temperature; SST is the sea surface temperature; DOS is the distance to the open sea; NSFD is the number of snow-free days. Outputs: Four multiple linear regressions, one for each year, were performed in turn to estimate a GDD value for each meteorological station. Because of the small sample size (9–13 stations), special attention was given to the indicators from which the quality of estimations could be appraised: test of significance of variables, R2 and root mean square error (RMSE), i.e. the difference between the estimated value (GDDest ) and the observed value (GDDref ). Upon completing each annual regression, the coefficient of regression associated with each explanatory variable (Eq. (3)) was known. A spatial model of some limited generality was derived from these annual models. Only data for the three years from 2012 to 2014 were used in the calculation because the year 2011 had no parameters for the GDD1 and DOS variables, hence, omitted from the regression. This mean model was calculated in two steps. The first step was calculating the mean of the three regression coefficients (CRmean ) for each of the five explanatory variables. This mean value for CR was then multiplied by the corresponding value for each explanatory variable read in i, j (Eq. (4)). The five values were then summed to estimate GDDmean-est for each of the

628

D. Joly et al. / Ecological Indicators 66 (2016) 623–631

Table 7 Parameters of the regressions (intercept and regression coefficients (r)); coefficient of determination (R2 ) and RMSE; parameter mean values. Intercept

2011 2012 2013 2014

Regression coefficients

14.253 25.154 −63.809 −303.758

Mean

R2

RMSE

GDD1

GHSPI

SST

DOS

NSFD

0.232 0.961 1.636

0.344 0.448 0.201 0.138

2.05 1.99 0.78 1.40

1.019 0.886 0.691

3.170 0.932 2.271 3.227

0.96 0.91 0.80 0.83

30.4 46.3 80.6 86.9

0.943

0.269

1.275

0.865

2.4

0.87

61

12 (2012) or 13 (2013 and 2014) stations:

4.3. Map of Svalbard GDDmean-est

GDDmean-est[i,j] = (GDD1i,j × CRmean[GDD1] )

GDDmean-est applied to the whole of Svalbard (Fig. 2) shows first the strongly decreasing values of GDD from lowlands to mountains according to the GDD1 and NSFD parameters. Second, the eastern and northern parts stand out as being considerably colder than the central part of Spitsbergen where GDD5 values up to 500 are common. Some spots with GDDmean values up to 600 occur in the inner part of the wide valleys bordering the Isfjorden area.

+ (GHSPIi,j × CRmean[GHSPI] ) + (SSTi,j × Cmean[SST] ) + (DOSi,j × CRmean[DOS] ) + (NSFDi,j × CRmean[NSFD] )

(4)

where CRmean is the mean (years 2012, 2013 and 2014) coefficient of regression. The second step was calculating the mean (GDDmean-est ) for the 12 or 13 GDDmean-est[i,j] . (GDDmean-est ) is then subtracted from GDD[ref] : Intercept = GDDmean-est − GDD[ref]

(5)

This spatial model of GDD was then applied to each pixel i, j of Svalbard. The resulting map shows the spatial variation of GDD with a spatial resolution of 100 m.

5. Discussion The method used to model GDD was based on standard statistical approaches and on common sense in order to evaluate variables assumed to influence temperature variations. For example, with reference to other areas, we know beyond doubt that the North Atlantic current has a prominent influence on climate and temperature in particular. The GDD map for the whole of Svalbard (Fig. 2) makes allowance for these constraints. However, the GDD calculation depends on explanatory variables, some of which are worth discussing.

4. Results 5.1. Missing data

4.1. Annual GDD model The intercept (y-value at the origin) and the regression coefficients for the four summers are given in Table 7. The regression coefficient associated with each explanatory variable varies from year to year. For example, for one unit of GDD1, GDDest rises by 0.23◦ for 2012 and by 1.6◦ in 2014. The same goes for SST which rises from 0.078◦ for 2013 to 0.203 for 2011: between an SST of 0◦ (sea temperature north and north-east of Svalbard) and 7◦ (NyÅlesund), the rise in GDDest is 14◦ . The other variables are more stable from a year to the other. The GDD estimation is of high quality; especially for 2011 and 2012 (R2 is up to 0.9). The RMSE value is low for 2011 and 2012 (<46) and higher in 2013 and 2014 (>80).

4.2. Mean GDD model (GDDmoy-est ) By applying the mean spatial model to the stations for which the annual GDD value is known its quality can be evaluated (Table 8). The R2 and RMSE values are similar to those obtained by regression (Table 6). The mean model for the three years 2012–2014 is also of good quality (R2 = 0.86) with an average RMSE value (71).

Table 8 R2 and RMSE of the values estimated by the mean model.

2

R RMSE

2011

2012

2013

2014

Mean 2012–2014

0.91 54

0.86 66

0.79 86

0.80 95

0.86 71

The quality control of data is quite important to discuss because the missing data in the datasets can partly be a problem concerning the robustness of the results. The calculation of GDD1 is based on the average of 460 daily temperature (92 days per summer × 5 summers from 2001 to 2005) recorded in 45 stations in the Kongsforden area. Considering that the recordings were made under extreme climatic conditions, missing data (35%) reach a satisfactory rate. This rate is higher at the end of winter (on average 40% in June) than at the end of the summer (28% in August) after the devices had been controlled. Missing data in the databases provided free of charge by the Norwegian Meteorological Institute (13 stations for three months and five years) have in contrast a relatively low importance. Indeed, these data are just used to control the quality of estimates by calculating the RMSE.

5.2. Colinearity between SST and NSFD The high colinearity between SST and NSFD might have prompted us to omit one or other of the variables. Because the colinearity indicators (tolerance and VIF) were good, we included them both. But to assess their contribution to the global explanation of the variance of GDDref , we calculated R2 and RMSE by eliminating them in turn. Although SST and NSFD have only about 30% of non-common variance, they both contribute to improving R2 and sometimes very clearly so for example in 2011 (Table 9). The RMSE moves in the same direction although in some instances (2012 and 2013) with very slightly lower values for five variables.

D. Joly et al. / Ecological Indicators 66 (2016) 623–631

629

Fig. 2. The final GDD model applied to the Svalbard archipelago (except for the island of Bjørnøya).

5.3. Stepwise introduction of explanatory variables Fig. 3 shows the improvement of r2 as new explanatory variables are included in the model. This shows that while almost all the variables are useful, however, some are more important than others. The rate of variance explained by GDD1, zero in 2011, is low for the other years (r2 = 0.3 to 0.46). The introduction of the

two regional variables GHSPI and SST significantly increases the r2 values which reach 0.89 for 2012 and 0.75 for the other summers. The inclusion of DOS and NSFD improves the model’s performance, especially for 2011, and less so for the other years because of the colinearity between the added variables and the previous ones. It can, however, be concluded that all the explanatory variables are contributing to account for the variance of GDD.

Table 9 R2 and RMSE values for three cases: five and four (NSFD and SST alternatively removed) explanatory variables; the response variable is the sum of summertime temperatures for 2011–2014; DOS is not included in the regressions for 2011. 5 variables R2 2011 2012 2013 2014 RMSE 2011 2012 2013 2014

0.90 0.91 0.80 0.83 35.9 46.3 80.6 85.8

4 var NSFD removed 0.74 0.90 0.76 0.78 81.1 45.8 82.9 94.2

4 var SST removed 0.85 0.85 0.79 0.82 62.1 56.2 77.6 86.9

Bold figures indicate the highest r2 and the lowest RMSE values.

Fig. 3. Stepwise improvement of r2 by the introduction of new variables; GDD1 and DOS are not included in the regression for 2011.

630

D. Joly et al. / Ecological Indicators 66 (2016) 623–631

because the mean of the correlation coefficients does not adjust GDDref as well as annual regressions. However, the advantage is that the model is a general one, and as such, it can be applied to years with little empirical data, such as 2011, for which regressions do not work well. 6. Conclusion

Fig. 4. Mean GDD (◦ C) for nine elevation ranges.

5.4. GDDest reduction with altitude Elevation plays a key role in the spatial variation of temperatures and plants in Svalbard (Virtanen, 1996; Körner, 2000; Bruun et al., 2006; Sundqvist et al., 2013) as in any other location worldwide. The variation of GDDest with elevation is taken into account by two variables: GDD1 and NSFD. The result obtained accounts for this constraint. GDDest clearly falls with elevation (Fig. 4). For 2011 and 2012, the mean GDDest for elevations below 50 m is 350; with higher elevations, the decline in GDDest is moderated with the result that the values remain positive even for the higher elevation bracket (>1000 m). In 2014, which was a cold season, it was just 300 and reached 0 at elevations of just 400 m. In 2013, a paradoxical situation occurred where GDDest on the strandflat was higher (400) than in the previous two years, but it was much lower at elevations of more than 400 m and the gap widened in proportion to elevation. It can also be noted that, consistent with the value of the parameter associated with GDD1 (Table 4), the decline of GDDest with elevation was faster for 2013 and 2014. 5.5. Difficulties in assessing the model The sparse availability of temperature records creates a problem for evaluating the model and this is probably the weak point of this study. Depending on the year, only 9–13 locations have long-term temperature records. The low number of weather stations made it problematic to compute the parameters of the regressions. Thus, the year 2011, with records from just nine stations, has only three explanatory variables significant at the 10% level. For the same reason it was also problematic to evaluate the results and it is not possible to document the performance of the model closely with only the present meteorological data available. This is especially true when evaluating GDD estimations with elevation. In the previous section, we discussed a probable fall in GDDest with elevation. However, it is not possible to evaluate the performance of the model as there are no weather station higher than 28 m.a.s.l. (Svalbard Airport) to provide reference data. The evaluation problems are also connected to large-scale variation and particularly the effect of the North Atlantic current and cloud fraction. These variables are relatively well represented by the reference temperature records, while the effect of “continentality” is better documented through the west to east distribution of meteorological stations from Isfjord Radio to Longyearbyen via Barentsburg (Førland et al., 1997). 5.6. Mean spatial model The difference between the r2 and RMSE values obtained after each annual regression and after applying the mean spatial model (r2 slightly lower, RMSE slightly higher, see Tables 7 and 8) arises

The aim of this work was to bring some objectivity and quantification to the assumed temperature gradients claimed to have an impact on plant growth- and distribution in Svalbard. By using a solely spatial and statistical modelling approach, we have produced a model of the distribution of growing day degrees, which is plausible and consistent with the general opinion about temperature variations in Svalbard. The final GDD model (Fig. 2) contains both small- and largescale variables. Being a model, it obviously deviates from reality and that the magnitude of deviations varies with location and time. The GDD-model is the first of its kind covering Svalbard and should be amenable to future improvements. The northern and eastern parts of the Svalbard archipelago especially need to be better covered by empirical climatic data. 1. The final GDD successfully models the growing season on the Svalbard archipelago according to observed pattern: a. GDD-values decline with higher elevation as expected; b. Spitsbergen’s west coast is markedly milder than the coastal areas and islands in the northern and eastern parts of the archipelago which are influenced by the North Atlantic current and cold water or ice bodies; c. Central Spitsbergen is less cold than the more oceanic (peripheral) areas because of continentality. 2. The statistical analyses show that elevation is the most significant variable for explaining variations in GDD. As the topography of Svalbard has rough relief, composed mainly of mountains, glacial valleys and strandflats, this variable creates large variations in GDD over just short distances. 3. Of the large-scale variables, the continentality gradient and snow-free period are the most influential. In our opinion, the present study is a significant step towards constructing an objective, reliable spatial model to enable us to compute thermal conditions (GDD) at a high spatial resolution (100 m) and will be the basis for upcoming research objectives. The model provides a suitable indicator that is particularly relevant for plant growth but also useful for herbivore studies and other ecological topics where primary production is relevant. References Bliss, L.C., 1971. Arctic and alpine plant life cycles. Annu. Rev. Ecol. Syst. 2, 405–438. Bolstad, P., 2012. GIS Fundamentals: A first Text on Geographic Information Systems, 4th edition. Eider Press, pp. 674. Bruun, H.H., Moen, J., Virtanen, R., Grytnes, J.A., Oksanen, L., Angerbjörn, A., 2006. Effects of altitude and topography on species richness of vascular plants, bryophytes and lichens in alpine communities. J. Veg. Sci. 17, 37–46, http:// dx.doi.org/10.1111/j.1654-1103.2006.tb02421.x. Callaghan, T.V., Press, M.C., Lee, J.A., Robinson, D.L., Anderson, C.W., 1999. Spatial and temporal variability in the responses of Arctic terrestrial ecosystems to environmental change. Polar Res. 18, 191–197, http://dx.doi.org/10.1111/j.1751-8369. 1999.tb00293.x. CAVM team, 2003. Circumpolar Arctic Vegetation Map. (1:7,500,000 scale), Conservation of Arctic Flora and Fauna (CAFF) Map No. 1. U.S. Fish and Wildlife Service, Anchorage, AK. Elvebakk, A., 1990. A new method for defining biogeographical zones in the Arctic. In: Kotlyakov, V.M., Sokolov, V.E. (Eds.), Arctic Research: Advances, Prospects. Proceedings of the Conference of Arctic, Nordic Countries on Coordination of Research in the Arctic. Academy of Sciences of the USSR Commission on Arctic Research, Moscow, pp. 175–186. Elvebakk, A., 1994. A survey of plant associations and alliances from Svalbard. J. Veg. Sci. 5, 791–801, http://dx.doi.org/10.2307/3236194.

D. Joly et al. / Ecological Indicators 66 (2016) 623–631 Førland, E.J., Hanssen-Bauer, I., Nordli, P.Ø., 1997. Climate statistics and longterm series of temperature and precipitation at Svalbard and Jan Mayen. Klima, Report 21797. Norwegian Meteorological Institute, Oslo. Førland, E.J., Benestad, R.E., Hanssen-Bauer, I., Haugen, J.E., Skaugen, T.E., 2011. Temperature and precipitation development at Svalbard 1900–2100. Adv. Meteorol. 1, 1–14, http://dx.doi.org/10.1155/2011/893790. Hair, J.F., Anderson, R.E., Tatham, R.L., Black, W.C., 2006. Multivariate Data Analysis, 6th ed. Pearson Education Inc., Upper Saddle River, NJ. IPCC, 2013. Climate change 2013: the physical science basis. In: Stocker, T.F., Qin, D., Plattner, G.-K., Tignor, M., Allen, S.K., Boschung, J., Nauels, A., Xia, Y., Bex, V., Midgley, P.M. (Eds.), Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge University Press, Cambridge, United Kingdom/New York, USA. Joly, D., 1994. Ambiances climatiques instantanées au Spitsberg; pour une approche méthodique par niveau d’échelle. ‘Instantaneous weather ambiances at Spitsbergen; methodological approach by scale’ Annales Littéraires de l’Université de Franche-Comté, no 529. Diffusion Les Belles Lettres, Paris. Joly, D., Brossard, T., 2007. Contribution of environmental factors to temperature distribution at different resolution levels on the forefield of the Loven Glaciers (Svalbard). Polar Rec. 43 (4), 353–359. Joly, D., Nilsen, L., Brossard, T., Elvebakk, A., 2010. Plants as bioindicators for temperature interpolation purposes: analyzing spatial correlation between botany based index of thermophily and integrated temperature characteristics. Ecol. Indic. 10, 990–998, http://dx.doi.org/10.1016/j.ecolind.2010.02.007. Joly, D., Bois, B., Zaksek, K., 2012. Rank-ordering of topographic variables correlated with temperature. Atmos. Clim. Sci. 2, 139–147, http://dx.doi.org/10.4236/acs. 2012.22015. Karlsen, S.R., 2003. A method using indicator plants to map local climatic variation in the Kangerlussuaq/Scoresby Sund area, East Greenland. J. Biogeogr. 30, 1469–1491, http://dx.doi.org/10.1046/j.1365-2699.2003.00942.x. Karlsen, S.R., Elvebakk, A., Johansen, B., 2005. A vegetation based method to map climatic variation in the arctic-boreal transition area of Finnmark, northeasternmost Norway. J. Biogeogr. 32, 1161–1186, http://dx.doi.org/10.1111/j. 1365-2699.2004.01199.x. Körner, C., 2000. Why are there global gradients in species richness? Mountains might hold the answer. Trends Ecol. Evol. 15, 513–514, http://dx.doi.org/10. 1016/s0169-5347(00)02004-8.

631

Lenoir, J., Gégout, J.C., Marquet, P.A., de Ruffray, P., Brisse, H., 2008. A significant upward shift in plant species optimum elevation during the 20th century. Science 320, 1768–1771, http://dx.doi.org/10.1126/science.1156831. Moen, A., 1998. National Atlas of Norway: Vegetation. Statens Kartverk, Hønefoss. Nilsen, L., Arnesen, G., Joly, D., Malnes, E., 2013a. Spatial modelling of arctic plant diversity. Biodiversity 14, 67–78, http://dx.doi.org/10.1080/14888386. 2012.717008. Nilsen, L., Joly, D., Elvebakk, A., Brossard, T., 2013b. Modeling spatial covariation of summer temperatures and bio-indicators in an arctic costal area. Clim. Res. 58, 1–13, http://dx.doi.org/10.3354/cr01173. Ørbæk, J.B., Hisdal, V., Svaasand, L.E., 1999. Radiation climate variability in Svalbard: surface and satellite observations. Polar Res. 18, 127–134 http://www. polarresearch.net/index.php/polar/article/download/6565/7398. Pan, Y., Jackson, R.T., 2008. Ethnic difference in the relationship between acute inflammation and serum ferritin in US adult males. Epidemiol. Infect. 136, 421–431. Rogerson, P.A., 2001. Statistical Methods for Geography. Sage, London. Sundqvist, M.K., Sanders, N.J., Wardle, D.A., 2013. Community and ecosystem responses to elevational gradients: processes, mechanisms, and insights for global change. Annu. Rev. Ecol. Evol. Syst. 44, 261–280, http://dx.doi.org/10. 1146/annurev-ecolsys-110512-135750. Thuiller, W., Lavorel, S., Midgley, G.F., Lavergne, S., Rebelo, A.G., 2004. Relating plant traits and species distributions along bioclimatic gradients for 88 Leucadendron species in the Cape Floristic Region. Ecology 85, 1688–1699, http://dx.doi.org/ 10.1890/03-0148. Tuhkanen, S., 1993. Treeline in relation to climate, with special reference to oceanic areas. In: Alden, J. (Ed.), Forest Development in Cold Climates. Plenum Press, New York, pp. 115–134. Virtanen, R., 1996. Middle-northern (oro) arctic vegetation of northern Europe: Community types and vegetation gradients. Arctic and oroarctic vegetation patterns. In: Northern Europe as a Consequence of Topography, Climate Bedrock Conditions and Grazing. Acta Universitatis Ouluensis. Oulu University Press, Oulu, pp. 282. Walker, D.A., Raynolds, M.K., Daniëls, F.J.A., Einarsson, E., Elvebakk, A., Gould, W.A., Katenin, A.E., Kholod, S.S., Markon, C.J., Melnikov, E.S., Moskalenko, N.G., Talbot, S.S., Yurtsev, B.A., other members of the CAVM TEAM, 2005. The circumpolar Arctic vegetation map. J. Veg. Sci. 16, 267–282.