Spatial variation and seasonal dynamics of leaf-area index in the arctic tundra-implications for linking ground observations and satellite images

Vegetation in the arctic tundra typically consists of a small-scale mosaic of plant communities, with species differing in growth forms, seasonality, and biogeochemical properties. Characterization of this variation is essential for understanding and modeling the functioning of the arctic tundra in global carbon cycling, as well as for evaluating the resolution requirements for remote sensing. Our objective was to quantify the seasonal development of the leaf-area index (LAI) and its variation among plant communities in the arctic tundra near Tiksi, coastal Siberia, consisting of graminoid, dwarf shrub, moss, and lichen vegetation. We measured the LAI in the field and used two very-high-spatial resolution multispectral satellite images (QuickBird and WorldView-2), acquired at different phenological stages, to predict landscape-scale patterns. We used the empirical relationships between the plant community-specific LAI and degree-day accumulation (0 °C threshold) and quantified the relationship between the LAI and satellite NDVI (normalized difference vegetation index). Due to the temporal difference between the field data and satellite images, the LAI was approximated for the imagery dates, using the empirical model. LAI explained variation in the NDVI values well (R2adj. 0.42–0.92). Of the plant functional types, the graminoid LAI showed the largest seasonal amplitudes and was the main cause of the varying spatial patterns of the NDVI and the related LAI between the two images. Our results illustrate how the short growing season, rapid development of the LAI, yearly climatic variation, and timing of the satellite data should be accounted for in matching imagery and field verification data in the Arctic region.


Introduction
Vegetation monitoring is a tool for detecting the impacts of climate change on the composition and phenology of arctic ecosystems. For example, satellite image series spanning several decades have already revealed the large-scale greening of the arctic, a consequence of the increased plant growth and spatial expansion of shrubs and trees (Stow et al 2004, Forbes et al 2010, Frost and Epstein 2014. Spatially extensive observations of vegetation are equally essential when vegetation parameters, such as the leaf-area index (LAI), are included in ecosystem models (e.g. Cramer et al 2001, Melton et al 2013. Many key properties of vegetation can be reasonably well inferred from spectral reflectance and thus mapped for large areas, using remotely sensed data (Laidler and Treitz 2003, Laidler et al 2008, Ustin and Gamon 2010. However, the temporal and spatial scales of currently available remote-sensing products may pose a challenge during mapping of spatially heterogeneous landscapes, such as the northern tundra (Virtanen and Ek 2014).
The pixel size of the commonly used satellite imageries, e.g. those obtained from the Landsat and Moderate Resolution Imaging Spectroradiometer (MODIS) satellites, ranges from tens to hundreds of meters and cannot reveal fine-scale heterogeneity in vegetation and ecosystem properties (Laidler et al 2008, Virtanen and Ek 2014, Mora et al 2015, Bratsch et al 2016. This complicates the examination of plant growth responses to warming, which may vary among neighboring communities (e.g. McManus et al 2012, Bratsch et al 2016). To resolve this problem, images of very high spatial resolution (VHSR, 0.3À2 m pixel size) have become available in recent years. Their usage, however, is hampered by the high price, limited temporal availability, which is caused by the relatively long revisit periods of high-resolution image sensor satellites in the same location, and by the frequent cloud cover and low solar angle in the Arctic (Hope and Stow 1996, Rees et al 2002, Stow et al 2004, Westergaard-Nielsen et al 2013. Therefore, the bestquality image often does not temporally match the ground-truth data.
This temporal mismatch is a source of uncertainty in the end product, because the reflectance is dependent on the amount of biomass, plant species composition, and the water, nutrient, and pigment contents of plant tissues, all of which are affected by the growth stage of the vegetation (e.g. Ustin and Gamon 2010). Therefore, the spectral responses to the seasonal changes in vegetation properties should be better understood (Garrigues et al 2008, Ustin and Gamon 2010, Rautiainen et al 2011, Westergaard-Nielsen et al 2013. Since the growing season is short and the changes in the LAI and biomass of the vegetation are rapid in the Arctic, it is likely that even a small time difference between the image and the ground-truth data can result in wide variations in terms of growth stage. In this study, therefore, our objective was to quantify the spatial and seasonal variation in LAI among the dominant plant communities in the arctic tundra and to evaluate how the seasonality of the vegetation affects the interpretation of vegetation structure from satellite images. We measured the seasonal development and spatial pattern of the LAI in the coastal arctic tundra near Tiksi, NE Russia, in the summer of 2014 and examined how the normalized difference vegetation index (NDVI), derived from the reflectance data, varied among the plant communities and between the VHSR multispectral satellite images of two different growing seasons. To quantify the dependence of the seasonal LAI development on the weather and for reconstructing the LAI values, we developed regression models between the LAI, degreedays (DD) accumulation, and satellite-based NDVI. Based on these models, we then evaluated the impacts that the temporal mismatch between the satellite imagery and field data may have on the interpretation of the NDVI and LAI distributions in the landscape.

Study site
The study site is located about 500 m from the coast of the Arctic Ocean near the Hydrometeorological Observatory of Tiksi in NE Russia (71.5936°N, 128.8850°E, figure 1). The climate at Tiksi is arctic, with very cold and windy winters, short but relatively warm summers, and short shoulder seasons between these. The mean annual temperature was À12.7°C and the mean annual precipitation 323 mm in The site represents a typical coastal tundra of Eastern Siberia with alkaline bedrock and high plant species diversity. We focused on an area of approximately 1 km 2 around the micrometeorological station, established in 2010 for eddy covariance (EC) measurements of the land-atmosphere exchange of water, heat, carbon dioxide (CO 2 ), and methane (CH 4 ) (Uttal et al 2016, figure 1). The terrain around the EC mast was relatively flat; in addition to microtopographic variation there was a gentle slope rising towards the north and a small stream running through the area. The vegetation and land-cover types within the area were classified, using ground-based visual judgement during an expedition in 2012. These land cover classes (LC) were characterized as (1) dry fen, (2) wet fen, (3) bog, (4) lichen tundra, (5) shrubmoss tundra, (6) graminoid tundra, (7) flood meadow, (8) bare ground, and (9) water.
The fen and bog were peat-forming environments, while the other land cover classes showed no clearly discernible peat. Sedges (Carex L. spp.) characterized the vascular plant vegetation in the fens (figure 2). Sphagnum mosses (Sphagnum L.) and feathermosses (e.g. Pleurozium schreberi (Willd. ex Brid.) Mitt) with shrubs were abundant in the dry fens, while the moss cover was sparse, due to aboveground water in the wet fens. The bogs showed typical microtopographic variation, and their vegetation was characterized by the presence of dwarf shrubs, dwarf birch (Betula nana L.), Sphagnum, and feathermosses. The vegetation of the flood meadows along the stream and drier graminoid tundra was dominated by graminoids (sedges, grasses) and willows (Salix L. spp.). Abundant feathermoss coverage on the ground layer and dwarf shrubs in the field layer characterized the shrub-moss tundra. Lichen tundra patches alternated with stony bare-ground surfaces.
Flow of the study Several steps were needed to obtain time series of LAI, to model LAI for each land cover class (LC) for the years of satellite data, and to produce the LAI maps over the study area. The steps were, in brief, as follows, while the details are given in following paragraphs.
1. Phenological dynamics of vegetation: Vegetation surveys (% cover and mean height of each plant functional types) in sample of plots seven times over the study period in 2014.
2. Estimation of LAI on basis of % cover and height: Sample of plots were harvested after vegetation survey during the peak biomass and LAI of harvested material was measured at PFT level. Data were used to develop regression models to predict LAI.
3. Producing time series of total and vascular plant LAI: A model using degree-day accumulation and chilling temperature accumulation as drivers was fitted for each land cover class and LAI was modelled for years 2005, 2012, and 2014. 4. Mapping spatial distribution of LAI using satellite data: LAI was modeled over the study area using relationship between satellite derived NDVI and LAI in the study plots.

Field data
The field data on the vegetation were collected in the summer of 2014, which had a warmer (871 DD) growing season than the long-term average. The 92 inventory plots with radius of 2.5 m were placed along 16 compass points at regular distances of 25, 50, 75, 100, 150, and 250 m from the EC mast (figure 1). Several additional plots were monitored at distances of 300, 350, and 400 m to balance the number of plots in the various LCs. In each plot, the vegetation was inventoried in four subplots (45 cm Â 45 cm in area), located 2 m from the plot midpoint in four main compass directions. Each plot was classified according Environ. Res. Lett. 12 (2017) 095002 to the a priori LC scheme described above (dry fen, wet fen, etc.), and the plot midpoint was georeferenced, using a Global Positioning System (GPS) device and a measuring tape to achieve a location accuracy of 1-3 m. The vegetation was surveyed to characterize the species composition and to quantify the LAI and its seasonal development in the various LCs of the tundra. The vegetation was inventoried as plant functional types (PFTs), following the typification by Hugelius et al (2011), which is a modification of that by Chapin et al (1996). The PFTs included: (1) Sphagnum mosses, (2) other mosses, (3) lichen (4) dwarf shrubs, (5) deciduous shrub Betula nana, (6) deciduous shrub Salix spp., (7) herbaceous species, and (8) graminoids.
The projection cover percentage of each PFT was visually estimated, and the mean height was measured in each of the subplots during the main survey, 23-24 July 2014. These values were then averaged for the main 2.5 m radius plot to relate the vegetation and the satellite image-based spectral reflectance. The seasonal development of the vegetation in one of the subplots was monitored by seven successive surveys performed between 2 July and 15 August 2014. Another subplot was harvested immediately after the main survey to quantify the one-sided LAI. The vascular plant material harvested was scanned, using a Canon MP Navigator EX scanner (Canon Inc., Tokyo, Japan) and by calculating the green surface area (= LAI) of the scanned images, using GNU Image Manipulation Program 2 (GIMP 2) software. The LAI of the mosses was estimated as a projection coverage, i.e. we determined that a 100% cover would represent an LAI of 1 (Riutta et al 2007). This approach underestimates the true multilayered leaf area of mosses, but probably estimates reasonably well their light-capturing and reflectance properties, due to the lower pigment and nutrient contents in moss tissues (Tieszen and Johnson 1968, Moore et al 2006, Street et al 2012. Using the data of the harvested subplots, we then calculated for each PFT the relationship between the areal cover, plant height, and LAI and, using these relationships, estimated the LAI for each subplot at the time of the main survey (table 1). These relationships were also used, assuming space-for-time substitution, to estimate the LAI for the subplots that were monitored for seasonal dynamics.

Satellite image acquisition and processing
To examine the spatial and temporal variation in the vegetation patterns within the study area, we acquired two VHSR multispectral satellite images from the archive of DigitalGlobe (Westminster, CO, USA). To enable the comparison of images taken under different atmospheric conditions, the images were corrected for atmospheric scattering and transformed into surface reflectance values, using the dark-object subtraction method (Chavez 1988).
The images, QuickBird (QB, DigitalGlobe, 15 July 2005) and WorldView-2 (WV2, DigitalGlobe, 12 August 2012), were chosen because they were of good quality and showed the best temporal matching with the collection of the peak season field verification data (23-24 July 2014) in terms of calendar days. The growing season of the QB image was shorter and cooler (10 June-30 September 2005, 646 DD) than the growing seasons of the WV2 image (22 May-30 September 2012, 1071 DD) and the verification data (6 June-24 September 2014, 863 DD). Thus, the QB image captured a period during which the vegetation was in the fast-growing phase, while the WV2 image captured the peak LAI period (figure 3 (a)À(c)).
The mean reflectance values were extracted for circular plots with a 2.5 m radius. The reflectance data were used to calculate the normalized difference vegetation index, NDVI = (NIR À VIS)/(NIR þ VIS), which describes the absorbance of the red portion of visible (VIS) light and the reflectance of near-infrared (NIR) radiation by green vegetation. Thus, the NDVI is an indicator of the quantity and photosynthetic capacity of green vegetation and has commonly been used in spatial extrapolations of LAI (Tucker 1979, Laidler and Treitz 2003, Shaver et al 2013. Supervised land-cover classification was carried out, based on a WV2 image (12 August 2012) to visualize the spatial distribution of vegetation in the area (figure 1).
Temporal modeling of the LAI We examined the factors determining the seasonal development of the LAI to estimate it for the specific dates of the two satellite images. Since our field data on Table 1. Regression models describing the dependence of the leaf-area index (LAI) of the various plant functional types on their areal cover (C, %) and height (H, cm) in the Tiksi field plots surveyed in 2014. All regressions and parameters were significant at p < 0.05.
In this equation, t is the time, S 1 is the DD accumulated from all daily mean air temperatures over 0°C, and S 2 is the DD accumulated during the latter part of the growing season (after 15 July) from the daily mean air temperatures above 0°C, but below 10°C and a, b 1 , b 2 , c 1 , and c 2 are the parameters to be estimated. The equation was fitted to the total vascular plant LAI data for each LC (table 2). The fits obtained were favorable for the period covered by field observations in 2014, and the values outside the measurement period (2 July-15 August) were not used in any further analysis.

Data analysis
A paired sample Wilcoxon signed-rank test was used to determine whether the vegetation classes showed significant seasonal differences in their NDVI signals; i.e. the NDVI QB (15 July 2005) and NDVI WV2 (12 August 2012) were paired for each PCT. To further illustrate the seasonal changes in the various plant communities, we calculated the difference in NDVI between the late-season and early-season image (NDVI WV2 ÀNDVI QB ). The relationships between the LAI (either measured in the 2014 field survey or estimated for the actual dates of the satellite images) and the NDVI derived from the QB and WV2 images were examined, using regression analysis. This relationship is commonly used to extrapolate the Lichen tundra Shrub tundra Graminoid tundra Flood meadow Figure 3. Accumulation of (a)À(c) the temperature sum above 0°C (degree-days, DD), (d)À(f) vascular plant leaf-area index (LAI), and (g)À(i) total LAI in the dominant plant community types for the years of the field data (2014), QuickBird (QB) image (2005), and WorldView-2 (WV2) image (2012). The vascular plant LAI was modeled based on seasonal accumulated temperature (DD) (table 2), while the non-vascular LAI was assumed to be constant and directly proportional to the areal cover. The vertical dotted lines indicate the timing of the satellite images and of the field data.

Results
At harvest time, the average vascular plant LAI was 0.55 across all the harvested plots. The highest values were found in the wet fen and flood meadow, both dominated by graminoids, while the lichen and bareground tundra showed the lowest values ( figure 2(a)). The moss LAI was highest in the dry fen and bog ( figure 2(b)). The total LAI did not differ greatly among the LCs, excluding the lichen and bare-ground tundra, due to the contrasting distribution of mosses and vascular plants in the communities.
The seasonal amplitudes of the vascular and total LAIs was largest in those communities with abundant graminoid vegetation, i.e. in the wet fen, graminoid tundra and flood meadow (figures 3(d)À(i). Our modeled LAI showed that the development of the vegetation was delayed in the cool growing season of 2005 (figures 2(e) and (h)), whereas in 2012 the LAI developed rapidly, due to the early start of the growing season and higher DD accumulation (figures 3(c), (f) and (i)). The yearly differences were pronounced in the graminoid-dominated communities and small in the moss-and evergreen-dominated communities (3(d)À(f)). The varying meteorological conditions during the years of the satellite images led to differences in the LAI that were larger than would be expected, based solely on the difference between the dates of the images.
The differences in the LC-specific NDVI values between the two satellite images were in line with the differences in vegetation development illustrated by the LAI. In the QB image, which represents the LAI in the growth phase ( figure 4(a)), the NDVI values were largest in the shrub-moss tundra and bog, whereas in the WV2 image, which represents the maximum LAI, the NDVI maxima were found in the wet fen, flood meadow, and graminoid tundra ( figure 4(b)). The median difference in the NDVI QB and NDVI WV2 Table 2. Fit statistics and parameter values of the vascular plant leaf-area index (LAI) model (equation (1)) for each land cover class. The model was fitted to the community mean LAI of 7 measurement days.  values was significantly different from zero for all the plant communities, except for the lichen tundra and bare ground, which showed equally low NDVI values in both images ( figure 4(b) and (c)). The NDVI differences were larger among the LCs in the WV2 image than in the QB image. The relationship between the NDVI and LAI showed a typical exponential shape with reasonably favorable adjusted coefficients of determination (R 2 adj = 0.42-0.92, p > 0.05). The regression parameters, however, varied depending on the satellite image used due to the different timing of the images in relation to the field data and growth stage ( figure 5(a) and (b)). The variation in total and vascular plant LAI was better explained by the NDVI WV2 than the NDVI QB , due to the growth stage in the WV2 image better resembling the growth stage in the field data (figure 3). To illustrate the spatial variation of the LAI in the study area at the time of the satellite images, we calculated the NDVI-LAI regressions, using values adjusted with the phenological model ( figure 5(c) and (d)). For vascular LAI the relationship was roughly agreeable with a spatially more representative transfer function in Shaver et al (2007Shaver et al ( , 2013 based on multiple arctic sites ( figure 5(d)). In the scale of whole focus area particularly the vascular LAI differed between the images. In the WV2 image, the average vascular LAI was 30% higher and the graminoid LAI in the graminoid communities was 2.3-2.9 f old compared to their respective values in the QB image ( figure 6).

Discussion
In this study, we documented the spatial and temporal patterns of the LAI in the low-growth vegetation of the coastal arctic tundra and generalized these observations for the entire study area, using VHSR satellite images. Our study illustrates the impact of the spatial and temporal variation in the seasonality of vegetation, measured as the LAI, on the inference of remotesensing products in the tundra. Interestingly, both the plant communities and the LAI showed substantial small-scale spatial variation in our study area, which appears to be typical of tundra landscapes (Marushchak et al 2013, Virtanen and Ek 2014). The temporal dynamics of the LAI also differed among the plant communities. The graminoid-dominated vegetation showed intensive growth within the short arctic summer, leading to a large seasonal amplitude in the LAI. In the Carex-dominated wet fens and flood meadows, the sparse moss cover further amplified these seasonal differences in total greenness. The mixed vegetation of the evergreen and deciduous dwarf shrubs showed a smaller vascular plant LAI and a more abundant moss cover and thus less seasonal variation in its NDVI (figures 2 and 3). These differences in the LAI and NDVI dynamics were driven by the variation in the overwintering green biomass and leaf production dynamics among the various PFTs. Graminoids have little overwintering green biomass, while the evergreen dwarf shrubs show rapid Environ. Res. Lett. 12 (2017) 095002 green-up in the spring and gradual leaf turnover during the late summer (Johnson and Tieszen 1976, Saarinen 1998, Maanavilja et al 2010. Mosses, in turn, show little seasonal variation in nutrient and pigment contents (Moore et al 2006, Street et al 2012, but their water content affects their reflective properties (Vogelmann and Moss 1993).
We applied a simple regression model to estimate the interannual variation in the seasonal LAI development and to match the phenological phases of the LAI illustrated in the two satellite images. The images were taken in different years and were À9 and þ19 ordinal days apart from the date of the field data collection, which in turn represented the peak LAI season (figure 3). Both the NDVI values extracted for the survey plots and the modeled LAI values indicated that the earlier, i.e. À9 days, image represented a phase of vegetation development in which the graminoid LAI had not yet reached its growing season maximum (figures 3 and 4). Apparently, despite a larger deviation in the number of ordinal days, the later image was similar to the field data in terms of the growth phase.
We acknowledge modeling the LAI values for years with no field measurements, but the continuous data of daily GP max that we have available for 2010-2014 (S1) and the robust correlation between the LAI and GP max (e.g. Laurila et al 2001, however, provide strong support for the function that we chose for modeling the seasonal LAI patterns. Our simple model with DD as an environmental driver seemed to be suitable for our site and for this time period, but without cross-site validation it cannot be considered general, but rather a technical tool for local LAI estimation. Degree-day accumulation has found to be a reasonable driver of LAI in Arctic communities (e.g. Hollister et al 2005) and our model also included the effects of chilling temperatures in the latter part of the growing season. The phenological dynamics are also affected by the timing of thawing, thaw depth, previous year's conditions, soil moisture variations, herbivory, and photoperiod (Arft et al 1999, Marchand et al 2004, Körner and Basler 2010, Oberbauer et al 2013. These factors may be only partly accounted by the two-part DD model.
Our results show how the phenological stage of the vegetation can rapidly change in the short growing season of the Arctic and, therefore, that the timing of satellite image acquisition really matters. We illustrate how the spatial pattern of the LAI in the two images of the same area differ and how linking the vegetation patterns to the NDVI values of our QB image would result in a biased relationship and affect the interpretation of the remote-sensing products (Williams et al 2008, Ustin andGamon 2010). VHSR satellite imagery is a significant improvement over traditional imagery, such as Landsat, for mapping the vegetation and associated processes of heterogeneous landscapes, such as tundra (Laidler and Treitz 2003, Virtanen and Ek 2014, Siewert et al 2015, Shrestha et al 2016. However, it is not unusual for these images to be acquired in different years or dates than the field data, because the availability of VHSR images of the Arctic is limited, due to infrequent satellite visits and unfavorable cloud conditions and zenith angles in these areas (Hope and Stow 1996, Rees et al 2002, Stow et al 2004, Westergaard-Nielsen et al 2013. Our data showt that one must be cautious in interpreting images without knowing the phenological stage of the most abundant plant species and communities at the time of imaging. Moreover, these data add knowledge on arctic vegetation patterns and satellite derived NDVI in a less studied region in eastern Siberia representing a coastal  lowland middle Arctic tundra landscape. In order to spatially extrapolate LAI over the study area we applied an empirical relationship between plot scale LAI and VHRS NDVI with a typical exponential form (figure 5) observed across the Arctic (Van Wijk and Williams 2005, Stelzer and Welker 2006, Stoy et al 2009, Stoy and Quaife 2015. We found the relationship roughly similar between our study with VHSR derived NDVI and, for example, study by Shaver et al (2013) with multi-site data of fieldspectrometer derived NDVI ( figure 5(d)). However, similarities or differences in the NDVI-LAI relationship among sites and sensors deserve further attention, because it is commonly used for large scale extrapolations, but are known to be affected by changes in spatial scales , Stoy et al 2009, Stoy and Quaife 2015, between vegetation type or content (Stelzer andWelker 2006, Street et al 2007) and growing season . To our knowledge none of earlier studies has focused on similar Siberian calcareous tundra landscapes nor employed VHSR satellite imagery to establish LAI-NDVI relationship there. Data from multiple sites and including seasonal and multi-year variation would be valuable as verification material for developing models to generate LAI maps of the Arctic based on satellite images

Conclusions
Our objective was to determine the composition, distribution, and seasonal dynamics of the LAI in the plant communities of the arctic tundra and to assess how the seasonality of plant growth affects the interpretation of the vegetation signal derived from satellite images. Our observations have significant implications for the evaluation and planning of optical Earth observations and for compromising between spatial and temporal resolution. Understanding the small-scale spatial variation in plant communities, plant growth dynamics, and their constraints are highly important in heterogeneous landscapes when biological variables are interpreted, using VHSR Earth observation data. We conclude that the short growing season of high latitudes, in association with climatic variation, sets special requirements for linking nonmatching field data and satellite images in these areas.