A new NDVI measure that overcomes data sparsity in cloud-covered regions predicts annual variation in ground-based estimates of high arctic plant productivity

The study area comprised the central part of West Spitsbergen (Nordenskiöld Land peninsula) in the archipelago of Svalbard, located between approximately 77.7°N−78.4°N and 13.5°W−17.4°W (figure 1). The meteorological station at Svalbard Airport, located close to the archipelago's administrative centre Longyearbyen (figure 1(a)), recorded a mean July air temperature of 7.0 °C for 2000−2014 (Norwegian Meteorological Institute 2015); the annual mean air temperature over the same period was −3.5 °C. Various habitats occur within the study area, including heaths, fens and mires (Johansen et al 2012, Johansen and Tømmervik 2013). At higher elevations the vegetation is sparse, as shown in July NDVI values below 0.2 (figure 1(a)). Onset of the growing season, as defined by MODIS data and corresponding to time of flowering of the widespread polar willow (Salix polaris), normally occurs in mid to late June in the bottom of the valleys and in the first half of July at higher elevations (figure 1(b)), but there is high additional variability in space, as well as between years (Karlsen et al 2014).

Due to the short and intense period of plant growth in the High Arctic and the occurrence of frequent cloud cover (see Van der Wal and Hessen 2009), all possible cloud-free data are needed when determining the onset of the growing season from satellite-based NDVI data. Karlsen et al (2014) used MODIS Terra data (MOD09Q1 and MOD09A1 surface reflectance products, both with 8 day temporal composites), to map the onset of the growing season on Svalbard from 2000–2013 (figure 1(b)). To make use of as much cloud-free data as possible, a combination of three cloud removal methods (MODIS State Quality Assurance values, the authors' own algorithms and manual removal of cloudy parts of the image) were used (for more processing details, see Karlsen et al (2014)). In the current study, the MODIS-NDVI (232 × 232 m pixel size) dataset has been extended to include data from 2014. In addition, the dataset has been interpolated to daily time resolution, from which the annual onset of the growing season was calculated on a pixel-by-pixel level and defined as the time when the NDVI value for that year exceeded 70% of the 4 July to 3 August mean NDVI value from 2000−2014. Due to insufficient data in spring 2001, no granules of MODIS data were produced for the period 15 June−1 July 2001 (day-of-year 166−182). Since the spring 2001 data was also partly cloudy in the days prior to 15 June, the time gap was too long to allow interpolation by the adaptive Savitzky–Golay filter that we used. Therefore, we treated this period separately and used a simple linear interpolation method (weighted mean values from the cloud-free pixels before and after) to fill this gap (for spring 2001 only).

To validate the plant productivity proxy that we derived from MODIS-NDVI data (2000−2014), we used above-ground live vascular plant biomass data that was collected annually from 2000−2009 (as described in Van der Wal and Stien 2014). This plant biomass data covers marsh, ridge and heath vegetation types, and was collected in Semmeldalen (figure 1(a)). Field sampling was done annually around 1 August (range of medium sampling dates: 24 July−6 August), i.e. close to the estimated peak NDVI (see Results). For this study we have used annual mean biomass estimates from five sampling sites (100−1600 m apart) in the most common vegetation type, i.e. moist Luzula heath, which is also widespread across central Spitsbergen (Elvebakk 2005, Johansen et al 2012). Each sampling site was relatively homogeneous in terms of micro-topography and vascular plant community composition over at least 25 m². The biomass estimates are the sum of above-ground live vascular plant biomass, obtained through a combination of non-destructive shoot density counts in small (25 cm × 25 cm) permanent plots (10 per sampling site) and destructive shoot mass determinations outside these plots. Although the field sampling method allowed for the inclusion of small amounts of perennial above-ground biomass (from polar willow, of which 28% ± 7% (SD) of shoot tissue sampled was categorized as 'branch', and part of this will have been produced in years prior to when sampled), the biomass estimates largely consist of the annual growth of vascular plants (hence using peak plant biomass estimates as a field-based index of plant productivity).

To link the MODIS-NDVI data with the field-based plant productivity data two masks of different sizes were created. Firstly, in order to spatially match the (232 m pixel size) MODIS-NDVI data with the field biomass data, a single polygon enclosing moist Luzula heath vegetation (i.e. the target of the biomass sampling data used here) was delineated on a map in the field. This local mask represented densely vegetated areas with moist Luzula heath and hence excluded areas such as river, river gravel, snow patches, less vegetated ridges and wetlands. The resulting mask covered 7 MODIS pixels (0.38 km²), and is hereafter referred to as the local mask. Secondly, to test the NDVI field-data relationship for larger areas, a general mask was created (hereafter referred to as the regional mask), covering the entire study area depicted in figure 1. In this mask, only MODIS pixels with mean (2000−2014) July NDVI values above 0.5 were included (see dark green areas in figure 1(a)), resulting in a coverage of approximately 405 km². We used the NDVI threshold of 0.5 to ensure that the vegetation most representative of the field validation study area was included, and to reduce noise from sparsely vegetated areas.

Since several years in our study period were characterized by frequent cloud cover before and during the peak growing season (modal, median and average cloudiness of 6.5 recorded on an 8 point scale at Longyearbyen airport—see Van der Wal and Hessen 2009), we developed a new NDVI measure, which calculated the integrated NDVI from the onset (O) to the Peak (P) of the growing season (hereafter referred to as OP NDVI). For each year we calculated a cumulative measure of NDVI from time of onset of the growing season on a pixel-by-pixel level within the masks. However, for the peak of the growing season NDVI we used a fixed date, for all pixels for all years because time of maximum NDVI on a pixel-by-pixel and year-by-year basis showed several outliers. These were, again, likely due to the frequent cloud cover on Svalbard during the peak NDVI period and because some bryophyte-dominated areas may have their highest NDVI value either early or late in the season due to their lack of senescence (Arndal et al 2009). Hence, we estimated the average time of maximum NDVI across all pixels in all years, which for the Central Svalbard mask was 25 July (mean value for the 2000−2014 period). Therefore, to simplify data analysis and to reduce noise problems from cloud cover and bryophyte dominated areas we used 25 July (day-of-year 206) as time of maximum NDVI for both masks (local and regional) for all pixels in all years when calculating OP NDVI. For comparison of OP NDVI with more traditionally used NDVI products, we also determined peak NDVI values for both the local and regional mask areas using the unprocessed MOD09Q1 MODIS products.

Given the importance of summer temperature for high arctic plant growth (Van der Wal and Hessen 2009, Van der Wal & Stien 2014), and thus its likely influence on NDVI, we obtained daily mean temperatures from the meteorological station at Svalbard Airport near Longyearbyen (see figure 1(a) for its location). Two summer temperature indices were calculated for each year, i.e. average July temperature and total growing degree days (GDD). GDD (defined as the temperature sum > 0° C) were calculated from time of onset of the growing season (average for the respective mask each year) to average time of peak maximum NDVI (25 July) using daily mean temperatures.

Statistical analyses were carried out in R version 3.1.1 (R Development Core Team 2015). We used linear regression models to determine the relationship between (field-based) plant productivity and OP NDVI values derived from MODIS data for the two different masks (local and regional) for the period 2000−2009; the same analysis was run using maximum NDVI values instead of OP NDVI to contrast goodness of fit. We also used a similar model structure to analyse the extent to which between-year variation in OP NDVI (period 2000−2014) could be explained by the accumulated predictor variable GDD and, alternatively, onset of growth and July temperature, for the two different masks. The linear regression models used F-tests for analysis of variance, with the numerator and denominator degrees of freedom indicated by the subscript numbers.

Zoom In Zoom Out Reset image size

For the period with overlapping data for NDVI and plant biomass (2000−2009), average annual plant biomass in the moist Luzula heath plots varied between 30.0 and 44.8 g m⁻² (mean = 38.1, SD = 5.7, table A1), while the average MODIS based OP NDVI values for the time of onset of growth until the time of maximum NDVI varied between 14.6 and 21.8 (mean = 18.4, SD = 2.4) and 15.9 and 23.7 (mean = 19.9, SD = 2.5) for the local and regional masks, respectively. OP NDVI showed a strong positive relationship with the plant biomass data measured in the field (figure 2(a), figure 2(b); Pearson's correlation coefficients r = 0.75 and 0.71 for the local and regional mask, respectively), and the relationship was even stronger for the annual rates of change (figure A1). With the local mask, which surrounds the field data site, the OP NDVI increased by 6.62 for each 10 g m⁻² increase in plant biomass (F_1,8 = 10.2, p = 0.01, adjusted R² = 0.51; figure 2(a)). With the regional mask, where the mask covers all areas with high summer NDVI values, the OP NDVI increased by 9.70 for each 10 g m⁻² increase in plant biomass (F_1,8 = 8.08, P = 0.02, adjusted R² = 0.44; figure 2(b)). Compared with OP NDVI, maximum NDVI proved to be a poor proxy of peak plant biomass (figures 2(a) and (b), table 1). Using the local and the regional masks, the relationships between maximum NDVI and peak plant biomass were non-significant (F_1,8 = 3.08, p = 0.12 for the local mask and F_1,8 = 0.13, p = 0.73 for the regional mask). When calculating OP NDVI from the original MODIS product before any processing and cloud removal, correlation with the field data were then less than 0.3 for both masks (data not shown).

Analysis of the whole time-series (2000−14) revealed that OP NDVI was strongly and positively related to growing degree days (GDD) for both the local mask (F_1,13 = 44.9, p < 0.001; figure 2(e)) and the regional mask (F_1,13 = 41.5, p < 0.001; figure 2(f); table 1). GDD in turn was explained well by a multiple regression model including the additive effects (see table 1 for pairwise relationships) of the satellite imagery-based onset of spring and July temperature (F_2,12 = 27.9, p < 0.001, adjusted R² = 0.79). Accordingly, and unsurprisingly—due to its cumulative nature—OP NDVI was highly correlated with onset of growth (r = −0.89 and −0.93 for the local and regional masks, respectively; table 1). Onset of growth was also significantly correlated with field-based plant productivity (r = −0.65 and −0.62 for the local and regional masks, respectively; table 1), but not as strongly as with OP NDVI (table 1). Likewise, both OP NDVI and (in particular) field-based plant productivity were also positively correlated with July temperature (r = 0.69 (local), 0.65 (regional) and 0.94 (field-based), respectively; table 1). Thus, although careful interpretation is needed due to a tendency for collinearity, a multiple regression model (local mask: F_2,12 = 25.3, p < 0.001, adjusted R² = 0.78; regional mask: F_2,12 = 41.4, p < 0.001, adjusted R² = 0.85) indicated that OP NDVI values were effectively highest for years in which growth was estimated to start early and the summer was warm, which in turn explained the apparent positive effect of a high accumulated GDD (figure 2(e) and (d)). This pattern was also indicated in the shorter time-series (2000−09) of field-based productivity data (F_2,7 = 27.2, p < 0.001, adjusted R² = 0.85), although the effect of onset of growth (see table 1) was far from significant (p = 0.48 and 0.65 for local and regional mask, respectively) when fitted together with the strong July temperature effect (see also Van der Wal and Stien 2014).

High OP NDVI values, i.e. above ca. 20 units (corresponding to a plant biomass of more than 39 g m⁻² based on figures 2(a) and (b), were predominantly found at lower elevations of the main valleys of central Svalbard (figure 3). In two of the valleys (Adventdalen and Colesdalen, see figure 1(b)), OP NDVI values exceeded 25 in some spots, indicating an average peak plant biomass of more than 45 g m⁻² for these areas (given the temporal linear relationships in figure 2(a) and (b), and assuming that they can be extended). At higher elevations, the average OP NDVI values fell below 10 (i.e. < 25 g m⁻²).

Zoom In Zoom Out Reset image size

Large regional and inter-annual (2000−2014) variations in OP NDVI were apparent (figure 4), however there was no significant relationship between either the local or regional OP NDVI mask and year (r = −0.02 in both cases; table 1). In some years, OP NDVI was uniformly reduced across the whole of the study area, such as in 2008 with 84 % of the mapped area having an OP NDVI value of at least 2 units below the 2000−2014 average (figure 4). The years 2000 and 2009 also had lower than average OP NDVI values (58 % and 42 % of the mapped area > 2 units below average, respectively). Yet, the spatial pattern of reduced OP NDVI differed greatly between the two years, with reduced plant productivity in the eastern part of West Spitsbergen (including Sassendalen, Adventdalen and inner Reindalen) in 2000 and in the central part (including the valleys of Colesdalen, Semmeldalen and outer Reindalen) in 2009. Several years, such as 2002, 2006 and 2013, had generally greater than average OP NDVI (and, hence, plant productivity). In yet other years, some areas were more productive than average whilst other parts were less productive. This was, for example, the case for the years 2010 and 2014, but strikingly the area which had reduced productivity in 2010 (Reindalen) had greater than average productivity in 2014, highlighting the great variability in opportunities for vascular plant growth in different valley systems in different years. See appendix figure 1 for time-series plots showing the annual variation in OP NDVI, and other predictors (max NDVI, the onset of spring, GDD and July temperature), for the Nordenskiöld Land peninsula study area as a whole.

Zoom In Zoom Out Reset image size