You are on page 1of 12

Remote Sensing of Environment 114 (2010) 2719–2730

Contents lists available at ScienceDirect

Remote Sensing of Environment


j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / r s e

Annual changes in MODIS vegetation indices of Swedish coniferous forests in relation


to snow dynamics and tree phenology
A.M. Jönsson a,⁎, L. Eklundh a, M. Hellström a, L. Bärring a,b, P. Jönsson c
a
Department of Earth and Ecosystem Sciences, Division of Physical Geography and Ecosystem Analysis, Lund University, Sölvegatan 12, SE-223 62 Lund, Sweden
b
Rossby Centre, Swedish Meteorological and Hydrological Institute, SE-601 76 Norrköping, Sweden
c
Centre for Technology Studies, Environment and Society, Malmö University, 205 06 Malmö, Sweden

a r t i c l e i n f o a b s t r a c t

Article history: Remote sensing provides spatially and temporally continuous measures of forest reflectance, and vegetation
Received 12 February 2010 indices calculated from satellite data can be useful for monitoring climate change impacts on forest tree
Received in revised form 18 June 2010 phenology. Monitoring of evergreen coniferous forest is more difficult than monitoring of deciduous forest, as the
Accepted 18 June 2010
new buds only account for a small proportion of the green biomass, and the shoot elongation process is relatively
slow. In this study, we have analyzed data from 186 coniferous monitoring sites in Sweden covering boreal,
Keywords:
Coniferous forest
southern-boreal, and boreo-nemoral conditions. Our objective was to examine the possibility to track seasonal
MODIS changes in coniferous forests by time-series of MODIS eight-day vegetation indices, testing the coherence
NDVI between satellite monitored vegetation indices (VI) and temperature dependent phenology. The relationships
Norway spruce between two vegetation indices (NDVI and WDRVI) and four phenological indicators (length of snow season,
Phenology modeled onset of vegetation period, tree cold hardiness level and timing of budburst) were analyzed.
Scots pine The annual curves produced by two curve fitting methods for smoothening of seasonal changes in NDVI and
TIMESAT WDRVI were to a large extent characterized by the occurrence of snow, producing stable seasonal oscillations in
WDRVI
the northern part and irregular curves with less pronounced annual amplitude in the southern part of the country.
Measures based on threshold values of the VI-curves, commonly used for determining the timing of different
phenological phases, were not applicable for Swedish coniferous forests. Evergreen vegetation does not have a
sharp increase in greenness during spring, and the melting of snow can influence the vegetation indices at the
timing of budburst in boreal forests. However, the interannual variation in VI-values for specific eight-day periods
was correlated with the phenological indicators. This relation can be used for satellite monitoring of potential
climate change impacts on northern coniferous spring phenology.
© 2010 Elsevier Inc. All rights reserved.

1. Introduction and they may be hardened for a shorter period of time. De-hardening
and budburst may occur already in the beginning of the year when the
The recent global warming has induced earlier timing of spring seasonal temperature progression is slow and the risk of temperature
events (IPCC, 2007; Parmesan, 2006). In a meta-analysis with data backlashes with sudden frost episodes is high, which has raised
from Europe, the average advance of spring/summer was estimated to concerns of an increased risk for frost damage (Cannell & Smith, 1986;
be 2.5 days per decade during the period of 1971–2000 (Menzel et al., Jönsson et al., 2004; Morin et al., 2009).
2006). Changes in tree phenology may have implications for forest Monitoring of forest phenology is required to assess the impact of
management, in particular during the phase of forest regeneration ongoing climate change, which may vary among tree species due to
with selection of suitable tree species and provenances (Larsen, 1995; different sensitivity to seasonal changes in temperature, light and
Saxe et al., 2001). Leafing and flowering are visible signs of tree precipitation. Satellite remote sensing has the advantage of providing
response to seasonal changes, preceded by physiological processes spatially and temporally continuous measures of forest reflectance, and
associated with recovery from winter dormancy. Both the seasonal remotely sensed data are used in calculations of different vegetation
cold hardiness level (i.e. tolerance of freezing temperatures) and indices. The Normalized Difference Vegetation Index (NDVI) (Tucker,
timing of budburst are affected by the cumulative impact of ambient 1979) is one of the most widely used indices, and it has been applied to
temperature (Chuine, 2000; Kalberer et al., 2006). In a warming monitoring of spring phenological events over large areas (Beaubien &
climate, the trees may not develop the same cold hardiness as today, Hall-Beyer, 2003; Cleland et al., 2007; Schwartz et al., 2002; Stöckli &
Vidale, 2004; Zhang et al., 2006). NDVI is basically a measure of
‘greenness’, and it has been related to the absorption of photosynthet-
⁎ Corresponding author. Tel.: +46 46 222 94 10. ically active radiation (Grace et al., 2007) and GPP (Olofsson et al., 2008).
E-mail address: Anna_Maria.Jonsson@nateko.lu.se (A.M. Jönsson). A rather high correspondence between NDVI and leaf area index has

0034-4257/$ – see front matter © 2010 Elsevier Inc. All rights reserved.
doi:10.1016/j.rse.2010.06.005
2720 A.M. Jönsson et al. / Remote Sensing of Environment 114 (2010) 2719–2730

been found for Swedish coniferous forests (Tagesson et al., 2009), evergreen needleleaved trees. Further north, the evergreen needle-
though dense coniferous forests are commonly above the threshold for leaved trees dominate. Mid Sweden (latitude 58.7°N–63.5°N) is
recognizing asymptotic saturation effects (Eriksson, 2007). The same characterized by southern-boreal conditions, and North Sweden
phenomenon may occur during monitoring of agricultural areas, and the (latitude 63.5°N–68.3°N) by boreal conditions (Ahti et al., 1968).
Wide Dynamic Range Vegetation Index (WDRVI), derived from the Monitoring of soil conditions, understorey vegetation, crown condi-
NDVI, has been developed and used for reducing the sensitivity to tion and tree nutrient status (Table 1) is carried out at coniferous
saturation effects (Viña & Gitelson, 2005). forest sites by the Swedish forest agency according to the standards of
Satellite data have successfully been used for monitoring of birch ICP-forest (Haussmann & Lorenz, 2004). The size of the monitoring
phenology at high latitudes in Fennoscandia, using thresholds for the plots is about 50 m × 50 m. Only sites dominated by coniferous forest
onset of growing season (Karlsen et al., 2006, 2007, 2008). However, within a radius of 250 m according to Swedish CORINE land cover data
the phenology of evergreen tree species can be more difficult to (LMV, 2002) were included in this analysis (117 Norway spruce sites
monitor than the phenology of deciduous species. One reason is that and 69 Scots pine sites, Fig. 1). The forest site characteristics are
the annual changes in photosynthetic capacity of evergreen trees are influenced by climatic and geographical conditions showing latitudi-
less obvious, since green tissue is present all year round, and the new nal trends. An increase in latitude correlated with an increasing
shoots only account for a part of the green biomass. In a study of proportion of understorey vegetation, higher number of trees, older
boreal North America, NDVI of evergreen coniferous forest did not trees, lower productivity, less annual growth, and less nitrogen in the
reveal any consistent trend in relation to changes in temperature, soil (Table 2).
whereas increasing lengths of growing season were detected for the
nearby tundra region (Goetz et al., 2005). 2.2. MODIS data
A range of problems is encountered when monitoring forests at
high latitudes. The reliability and accuracy of winter values are In this study, satellite data products from the Moderate-Resolution
generally lower than during other seasons due to an influence of large Imaging Spectrometer (MODIS) instrument onboard the Terra
solar zenith angles, and to snow affecting the spectral reflectance satellite were used. Two MODIS tiles, labeled H18V02 and H18V03,
(Delbart et al., 2005; Schwartz, 1999). Smoothing methods for were used for extraction of data for the period of March 2000 to
outlining the seasonal patterns of vegetation are commonly used for August 2008. Gap filling was performed for the period A2001169
reducing effects dependent on view angle, errors in georeferencing using a copy of A2001161 for both tiles.
and low quality of data, though these methods may in turn add Vegetation indices (NDVI and WDRVI) were calculated from the
uncertainties related to the interpretation of data (Hird & McDermid, red and near-infrared reflectance using the eight-day 250 meter
2009). Several methods can be used for extracting phenological resolution surface reflectance (MOD09Q1) (Fig. 2), including the
measures from the annual curves (i.e. global thresholds, local thresh- reflectance data from bands 1 (red) and 2 (NIR) as well as the quality
olds, conceptual–mathematical functions and hybrid functions), and flags “band 1 data”, “band 2 data” and “MODLAND QA”. As the “cloud
the choice of method can have a large impact on the results as the state” quality flag included in the 250-m MOD09Q1 product was
methods differ in ability to track observed variability in timing of considered to be somewhat unreliable, the cloudiness information
phenology (White et al., 2009). “MOD35 cloud” from the eight-day 500 meter resolution surface
The objective of this study is to analyze the possibility to track reflectance (MOD09A1) was combined with the quality information
seasonal changes in coniferous forests by time-series of MODIS for the reflectance measurements “MODLAND QA bits” into an overall
eight-day vegetation index data, taking uncertainties of post- data quality indicator for each 250-m pixel used in the study.
processing into account by comparing the performance of two Similarly, the presence of snow and/or ice was derived directly from
vegetation indices smoothed by two fitting methods. The study was the 500-m data “MOD35 snow/ice”. (Note the concentration of pixels
carried out with the specific aims to 1) characterize the annual curves affected by snow and/or clouds during the winter season in Fig. 2).
of NDVI (most commonly used) and WDRVI (more recently The dates and orbits of the 250-m reflectance measurements were
developed for agricultural purposes, not yet thoroughly tested for assumed to be identical to the day-of-year value recorded in the
forest conditions) produced by the double logistic and the Savitzky– corresponding 500-m resolution dataset.
Golay fitting methods, and 2) analyze the coherence in between-year High quality was defined as occurring when the 250-m resolution
variations of vegetation indices and temperature dependent onset of reflectance bands were flagged as having ideal quality of all bands,
spring phenology. We used data from Swedish forest monitoring and the 500-m resolution “MOD35 cloud” flags had the value “clear”.
sites, covering three forest vegetation zones (boreo-nemoral, Reasonable quality corresponded to “less than ideal quality” for either
southern-boreal, and boreal conditions). Phenological indicators or both of the reflectance measurements and/or “mixed” cloud status.
were derived from simulations of coniferous cold hardiness, onset of All other combinations were classified as poor quality.
vegetation period, budburst and MODIS snow/ice flags.

2. Material and methods Table 1


Norway spruce and Scots pine sites monitored by the Swedish forest agency according
2.1. Monitoring sites to the standards of ICP-forest were included in this study. Monitored forest site
characteristics were used as explanatory variables in statistical analysis of vegetation
indices calculated from MODIS data.
In Sweden, almost 23 million ha are covered by forest. Norway
spruce (Picea abies) accounts for 42% of the standing volume, and Information category Site characteristics
Scots pine (Pinus sylvestris) for 38% (Loman, 2008). Norway spruce is Site position Latitude, longitude, height above sea level
commonly grown on more fertile soil than Scots pine. There is a Tree species Norway spruce/Scots pine
general south to north temperature trend across Sweden, with the Stand properties Number of trees, tree age, productivity class,
annual growth, standing volume 2004
average annual temperature during the period of 1961–1990 ranging
Soil nutrient status Carbon content, nitrogen content, base saturation
from 7 °C in the southernmost part to − 3 °C in the mountainous Soil properties Moisture class, pH (KCl), pH (H2O)
northernmost part. The length of growing season ranges from above Understorey vegetation % Cover
210 days in the south to less than 100 days in the north. South Sweden Crown defoliation % Needle loss, monitored during years 2000–2005
(latitude 55°N–58.7°N) is characterized by boreo-nemoral conditions Land cover class Coniferous forest, broadleaved forest, arable land,
other land use
with a mixture between summergreen broadleaved trees and
A.M. Jönsson et al. / Remote Sensing of Environment 114 (2010) 2719–2730 2721

Fig. 1. Map showing the distribution of coniferous forest monitoring sites in Sweden. Gray dots indicate Norway spruce sites (n = 117) and black dots Scotch pine sites (n = 69). The
two black lines show the borders between the North (latitude 63.5°N–68.3°N), Mid (latitude 58.7°N–63.5°N), and South (latitude 55°N–58.7°N), parts of Sweden.

As an alternative to the commonly used NDVI index, the Wide where the weighting coefficient (α) serves to attenuate the contri-
Dynamic Range Vegetation Index WDRVI has been proposed (Viña & bution of the NIR wavelength band for moderate-to-high green
Gitelson, 2005): biomass, hence reducing the saturation behavior inherent to the NDVI
metric as leaf area index increases. In the present study the value
α = 0.2 was chosen, based on experiences from a study of pine stands
WDRVI = ððα + 1Þ × NDVI + ðα−1ÞÞ  ððα−1Þ × NDVI + ðα + 1ÞÞ in southern Norway (Eklundh et al., 2009).

2.3. Phenological indicators


Table 2
Correlation between forest site characteristics (Table 1) and latitude, influencing climatic Large-scale field observations of coniferous phenology have not been
and geographical conditions, calculated for Norway spruce and Scots pine, respectively. carried out in Sweden, and we therefore used three temperature
Correlation coefficients are only given for significant correlations (pb 0.05).
dependent indicators to assess the variation between sites and years;
Site characteristics Norway spruce Scots pine cold hardiness level, onset of vegetation period, and the subsequent
r-values (n = 122) r-values (n = 72)
timing of budburst. Two phenological models, described in the following
paragraphs, were run with daily temperature data for the period of
Understorey vegetation 0.40 0.34
2000–2006. Gridded mean and minimum temperatures at 2 meter
Height above sea level 0.23 0.45
Number of trees 0.22 0.32 above ground (T2mean and T2min) were obtained from the E-Obs
Tree age 0.28 0.28 gridded dataset version 1.0 (daily temporal resolution, 0.22 degree
Productivity class − 0.63 Not significant spatial resolution), available via the EU-FP6 project ENSEMBLES
Annual growth − 0.43 − 0.49
(Haylock et al., 2008).
Standing volume − 0.54 − 0.48
pH (KCl) 0.41 0.31 The seasonal development of tree tolerance to frost was expressed
pH (H2O) 0.57 0.29 as the cold hardiness level, simulated with a daily time step (Jönsson
Base saturation 0.50 0.26 et al., 2004). For each time step, the daily mean temperature was used
Soil carbon content − 0.37 − 0.38 for calculating a steady-state hardiness level. The steady-state
Soil nitrogen content − 0.38 − 0.45
hardiness level corresponds to a threshold of frost tolerance induced
2722 A.M. Jönsson et al. / Remote Sensing of Environment 114 (2010) 2719–2730

Fig. 2. Data for two sites with evergreen coniferous forests; one with stable seasonal oscillations (A) situated at Lat 60.88°N, Lon 13.8°E and one with irregular and less pronounced annual
amplitudes (B) situated at Lat 56.99°N, Lon 15.49°E. The two upper subplots in A and B show NDVI data, processed with the software TIMESAT using the double logistic (solid black line) and
Savitzky–Golay (dashed gray line) methods. Unprocessed data are indicated as circles (○), high quality observations are indicated as circles with a dot (⨀), and a star (✰) indicates
observations influenced by snow or ice. The lower subplots show red reflectance data (red circles) and NIR reflectance data (black circles), used for calculating the NDVI (detailed data for
years 2003–2005). Filled circles indicate observations influenced by snow or ice. Observations influenced by cloud according to the 500-m MOD35 cloud flag are marked with a red and
black × (for red and NIR, respectively). Blue dots indicate the corresponding NDVI values (high quality only).

by temperature dependent chemical reactions and cellular adjust- hardiness level was above the steady-state hardiness level, it was
ments to withstand freezing temperatures. The minimum steady- adjusted according to a daily, temperature dependent, hardening rate.
state hardiness level was set to − 2 °C (typical value during summer, If the cold hardiness level was below the steady-state hardiness level,
indicating that frost damage can occur at −2 °C), and the maximum the cold hardiness level was adjusted according to a temperature
steady-state hardiness level was set to −50 °C (typical value during dependent de-hardening rate. However, de-hardening was not
cold winters). For each time step, the steady-state hardiness level was calculated between the autumn equinox (day 260) and the winter
compared with the cold hardiness level of the day before. If the cold solstice (day 355), accounting for winter dormancy. For comparison
A.M. Jönsson et al. / Remote Sensing of Environment 114 (2010) 2719–2730 2723

with satellite data, the cold hardiness level for the satellite budburst. The vegetation index at the timing of budburst corre-
observation day within each eight-day period was used. sponded roughly to 90% of the summer maximum, and this threshold
In this study, the onset of coniferous vegetation processes was was used for comparing the modeled timing of budburst with the first
calculated to occur after four consecutive days with a temperature eight-day period in spring when this level was reached.
above the threshold of 5 °C (Persson et al., 2007; Suni et al., 2003). The A rank correlation test was performed for each eight-day period to
accumulation of temperature impact on developmental processes analyze the within-site, between-year variation in vegetation indices
leading to budburst was counted from this onset. The budburst was in comparison with the onset of vegetation period, timing of budburst,
modeled to occur when 120 or 170 degree days had accumulated. The number of snow-indicated periods during winter and modeled cold
thresholds correspond to values reported for Norway spruce adapted hardiness level. The rank correlation tests for association between two
to the climate north and south of the boreal border at latitude 61°N variables by comparing their relative performance (rank), and the test
(Hannerz, 1994, 1999; Jönsson et al., 2004). In this study we used the can be used without assuming normal distribution or linear relation-
same thresholds for Scots pine as for Norway spruce for obtaining the ships (Sokal & Rohlf, 1995). By using the rank correlation comparison,
vegetation index value before budburst, as the Scots Pine has the between-year variation in temperature was captured without
approximately the same or somewhat higher thermal requirement being sensitive to the effect of local temperature conditions deviating
(Beuker, 1994; Rötzer et al., 2004). from grid cell averages and provenance-specific phenological timing,
thereby reducing uncertainties associated with phenological model-
2.4. Analysis of data ing and quality of climate data.
All modeling and data analysis were performed with MATLAB 7.0.4
MODIS data were processed with the software TIMESAT (Jönsson (The MathWorks Inc.).
& Eklundh, 2002, 2004), ver. 2.3, that has been developed for fitting
mathematical functions to time-series of satellite data in order to 3. Results
outline the seasonal progression of vegetation development. The
analysis was based on data from 2001 to 2006, i.e. the years with 3.1. Characterization of NDVI and WDRVI annual curves
complete MODIS and E-Obs data sets. Data from year 2000 was used
for initiating curve fitting and cold hardiness modeling. Two model The curves produced by TIMESAT smoothing of vegetation indices
fits were used, double logistic and Savitzky–Golay. The double logistic were characterized by stable seasonal oscillations for some sites,
function has been found to preserve the NDVI signal integrity (Hird & whereas irregular and less pronounced annual amplitudes were
McDermid, 2009) whereas Savitzky–Golay filtering using a relatively observed for other sites. Most of the sites were in-between these two
narrow window size is able to capture rapid changes by local fitting, extremes (Fig. 2). However, there is a certain geographical differen-
but is more sensitive to noise. The data were fitted using upper tiation in that irregular curves with less pronounced annual amplitude
envelope-weighted least-squares methods, assuming negative biased were more common in the South and Mid Sweden than in the North
noise, and data were weighted according to the quality indicators (Fig. 3A). Furthermore, the between-year variation in annual
ensuring that cloudy or otherwise low-quality pixels had minimum amplitude was typically more pronounced in the South than in the
influence on the fitting (Fig. 2). The high quality weight was set to 1, North, but this regional differentiation was sensitive to curve fitting
the less than ideal quality to 0.5, and poor quality to 0. procedure as it was more emphasized by the double logistic method
The three vegetation zones were characterized by analyzing than by the Savitzky–Golay method (Fig. 3B). For all three geograph-
variations in annual amplitude of NDVI and WDRVI, processed by ical regions, the variation in annual amplitude was mainly attributed
TIMESAT using the two smoothing methods. The between-year to variation during winter. The high variation during winter was in
variation was analyzed by calculating the standard deviations of turn caused by to two factors, data quality (including cloud) (Fig. 4)
annual amplitude, winter minimum and summer maximum for each and occurrence of snow (Fig. 5). The data quality was lower during
monitoring site. The results were expressed as the mean value for winter than during other seasons, and particularly low in the North.
each vegetation zone. Whereas low-quality data were not included in the analysis, the curve
Regression analyses were carried out to evaluate the influence of fitting procedures were indirectly affected by missing data.
site characteristics (Table 1) on TIMESAT processed vegetation For all sites, the average vegetation index value during periods
indices. The summer maximum vegetation index was used for with snow was below 0.5 for NDVI and below − 0.2 for WDRVI.
evaluation, as this was the most stable measure (Fig. 3). Site averages However, for all sites the highest values with snow were larger than
over years were calculated to get a robust measure of the dependent the lowest values without snow. This overlap was partly related to
variable, and site characteristics were used as explanatory factors. seasonality, i.e. vegetation indices measured during brief periods
The original MODIS data (i.e. not processed by TIMESAT) were with snow cover late in winter season (March–April) had generally
analyzed in order to investigate causes behind observed variations in higher values than vegetation indices measured during midwinter
annual amplitude. The proportion of high quality records for each (Fig. 6A, B). This was related to an increase in both red and NIR
eight-day period was calculated for all three vegetation zones in order reflectance during snow periods occurring later in winter season
to assess seasonal variations in data quality. The influence of snow (Fig. 2). Within each eight-day period, vegetation indices for years
was analyzed in a similar way by comparing observations with and with snow were lower than for years without snow.
without snow as indicated by the MODIS snow/ice flags. In addition, The within-site variation in vegetation indices was much less
vegetation index values for periods with and without snow were pronounced during summer than during winter (Fig. 3B). The summer
compared. The length of snow season was calculated as the sum of maximum of the annual curves captured between-site variations in
eight-day periods indicated by the MODIS snow/ice flag as influenced forest properties, being positively correlated with annual growth and
by snow during the first half of the year (periods 1 to 23). standing volume, and negatively correlated with crown defoliation
In search of threshold values generally applicable for assessing (Table 3).
spring phenology, the TIMESAT processed vegetation indices were
analyzed in relation to the modeled timing of budburst. To assess the 3.2. Vegetation indices in relation to temperature dependent phenology
influence of the new shoots on the vegetation indices, a ‘green
amplitude’ was calculated for all sites and years. The green amplitude The timing of phenological events was mainly determined by the
was defined as the difference between the summer maximum index latitude and further modulated by between-year variations. Onset of
value and the index value for the eight-day period just before vegetation processes occurred in January to June (eight-day periods
2724 A.M. Jönsson et al. / Remote Sensing of Environment 114 (2010) 2719–2730

Fig. 3. A) Mean vegetation index value and B) between-year variations, expressed as standard deviation, in winter minimum, summer maximum and annual amplitude for two
vegetation indices (NDVI and WDRVI) processed by two curve fitting methods, double logistic (DL) and Savitzky–Golay (SG). The standard deviations are site mean values,
calculated for all monitoring sites (n = 186) using data from 2001 to 2006.

Fig. 4. Quality of MODIS data during 2001–2006 with 46 eight-day periods per year. The boxes show the median, upper and lower quartile of the fraction of high quality observations
in the satellite monitoring data set covering 186 sites dominated by coniferous forest in North (n = 14), Mid (n = 88) and South Sweden (n = 92). The whiskers show the 1.5
interquartile range, and outliers are indicated by +.
A.M. Jönsson et al. / Remote Sensing of Environment 114 (2010) 2719–2730 2725

Fig. 5. Occurrence of snow during 2001–2006 within the 46 eight-day periods per year. The boxes show the median, upper and lower quartile of the fraction of snow/ice observations
in the satellite monitoring data set covering 186 sites dominated by coniferous forest in North (n = 14), Mid (n = 88) and South Sweden (n = 92). The whiskers show the 1.5
interquartile range, and outliers are indicated by +.

Fig. 6. A) Unprocessed NDVI values for all sites (only high quality data, sorted in latitudinal order with data from the northernmost site at the top of the figures and data from
the southernmost site at the bottom of the figure) and eight-day periods not influenced by snow according to the MODIS snow/ice flag. B) Unprocessed NDVI values for all sites and
eight-day periods influenced by snow according to the MODIS snow/ice flag.
2726 A.M. Jönsson et al. / Remote Sensing of Environment 114 (2010) 2719–2730

Table 3 used for assessing the timing of budburst as it varied considerably


Correlation coefficients between forest site characteristics (Table 1) and summer between sites and years, commonly occurring much earlier in season
maxima obtained for two vegetation indices (NDVI and WDRVI) as processed by two
curve fitting methods (double logistic and Savitzky–Golay). Norway spruce (n = 117)
than the timing of budburst (Fig. 7B). Sites with a pronounced annual
and Scots pine sites (n = 69) were treated separately. Only statistically significant amplitude, however, displayed a substantial increase in vegetation
correlations (p b 0.05) are shown. indices between the onset of the vegetation period and the timing of
budburst (illustrated for one site in Fig. 8). This is related to the
Tree Measure NDVI WDRVI
species melting of snow, and in the North the melting of snow can lead to a
Double Savitzky– Double Savitzky–
substantial increase in the vegetation indices also after budburst. This
logistic Golay logistic Golay
(r-values) (r-values) (r-values) (r-values) was particularly pronounced for Scots pine stands, showing higher
green amplitudes in the North than in the Mid and South Sweden
Spruce Annual growth 0.36 0.29 0.34 0.32
Standing volume 0.28 0.25 0.26 0.25
(Fig. 9). The Savitzky–Golay method that is sensitive to local variations
Crown defoliation − 0.20 − 0.16 − 0.23 − 0.20 in index values produced more pronounced green amplitudes than the
Pine Annual growth 0.64 0.58 0.64 0.59 double logistic method.
Standing volume 0.56 0.53 0.56 0.53 Whereas it was not possible to determine the annual timing of
Crown defoliation − 0.29 − 0.30 − 0.30 − 0.28
budburst using vegetation index thresholds, the interannual variabil-
ity of vegetation indices within each eight-day period covaried with
1–20), end of de-hardening in March to June (periods 16–22), timing the interannual variability of the four temperature dependent
of budburst in May to June (periods 17–23), and start of hardening measures of spring progress (Fig. 10). Years with an early onset of
in July to December (periods 23–45). The average NDVI value at vegetation processes and early timing of budburst were associated
the timing of budburst (Fig. 7A) corresponded approximately to with high vegetation index values, whereas years with late timing
90% of the site-specific summer maximum value (cf Fig. 9). The first were associated with low index values (i.e. negative rank correlation).
eight-day period when the 90% level is reached could however not be Longer snow seasons with late onset of spring were associated with
low vegetation indices, and shorter snow seasons with high
vegetation indices. On the other hand, deeper levels of cold hardiness
were associated with low vegetation indices, and lower levels of cold
hardiness with high vegetation indices (i.e. positive rank correlation).
The correlations were not statistically significant for all sites and
monitoring periods, due to uncertainties associated with data quality.
Nevertheless, the trend over adjacent periods indicated that the
strength of the correlations were dependent on the climate of the
geographical regions, and this can be used for monitoring interannual
variations in onset of spring.
The four indicators of spring processes produced complementary
assessments, with the relative importance during different eight-day
periods being related to indicator-specific timing and interannual
variation (Fig. 10). The length of the periods with correlations for the
four phenological indicators were partly overlapping, due to the
cumulative impact of temperature on snow melting and timing of
budburst. The vegetation indices were therefore correlated with the
length of snow season also in periods later than the latest snow record
(Fig. 5), and vice versa for the correlation between vegetation indices
and the timing of budburst. The regional-specific trends were
somewhat more pronounced for WDRVI than NDVI, but without any
major difference between the two smoothing methods.

4. Discussion

Satellite monitoring of forest phenology requires a defined


relationship between tree phenology and vegetation indices, and
there are a wide range of methods for estimating the start of spring
using remote sensing data, most of them producing estimates that are
related to or influenced by the snow dynamics (White et al., 2009). In
this study we analyzed time-series of satellite data for northern
coniferous forests, covering a latitudinal gradient in duration of snow
cover and commencement of spring phenology. We found some
support for relating the vegetation indices to temperature dependent
processes in spring. It was, however, not possible to establish any
absolute or relative threshold values for assessing the timing of
budburst, as the general outline of the annual cycle was more related
Fig. 7. A) The timing of budburst in relation to the eight-day periods for 2001–2006. to snow dynamics than to small and relatively slow changes in needle
B) The first period in spring when NDVI exceeded 90% of the summer maximum value. biomass. Neither can analysis of the rate of change in vegetation index
Data for NDVI were processed with the software TIMESAT using the double logistic between eight-day periods be used for monitoring of budburst in
method. The boxes show the median, upper and lower quartile for 186 monitoring sites
with Norway spruce and Scots pine. The whiskers show the 1.5 interquartile range, and
Sweden, due to the interference between snow melting and greening-
outliers are indicated by +. (See Figs. 2 and 8 for examples of site-specific annual up in combination with uncertainties associated with single vegetation
curves). index values.
A.M. Jönsson et al. / Remote Sensing of Environment 114 (2010) 2719–2730 2727

Fig. 8. Data for NDVI and WDRVI, processed with the software TIMESAT using the double logistic (DL) and Savitzky–Golay (SG) methods for one site with Norway spruce, situated at
Lat 64.24°N, Lon 17.39°E. The curves represent one year each (2001–2006), and modeled onset of vegetation processes (⨀) and timing of budburst (*) are indicated. Note the
interannual variation in vegetation indices for the two phenological indicators.

Snow alters surface reflectance, and snow covered forests may likely to be affected by snow (Suzuki et al., 2003). This study indicated
have very low NDVI values (Myneni et al., 1997). Winter-time that a fixed threshold value would be of little use for monitoring of
monitoring conditions are therefore very variable and the reflectance coniferous forests at northern latitudes due to the large overlap
data are not related to forest properties per se. The between-year between records with and without snow. The snow/ice flags may
variation in snow cover had a large influence on the annual amplitude provide a better option for sorting out records influenced by snow.
in South and Mid Sweden, whereas the variation was less pronounced Green coniferous canopies can, however, be observed even during
in the North due to longer periods with snow cover and a less variable winters with snow covered ground (Suzuki et al., 2001), as the snow
seasonal progression. It has been suggested that vegetation properties on branches is likely to disappear due to wind and sun exposure
should only be examined at NDVI values above 0.2 as they are less whereas the ground snow is more likely to stay for longer periods.

Fig. 9. ‘Green amplitude’, the difference between the VI value at the period before budburst and the summer maximum.
2728 A.M. Jönsson et al. / Remote Sensing of Environment 114 (2010) 2719–2730

Fig. 10. Rank correlation analysis of the interannual variation in vegetation index values and the temperature dependent onset of vegetation processes, timing of budburst, length of
snow season, and modeled cold hardiness level. The correlation coefficient (r-value), calculated for each site and eight-day period from January to June (i.e. periods 1–23) using data
from 2001 to 2006, expresses the degree of association. The r-values represent the mean across sites in the North (n = 14), Mid (n = 88) and South (n = 92) Sweden. Data shown for
WDRVI processed with TIMESAT using the Savitzky–Golay method.

This creates a need for evaluating the performance of the snow/ice by remote sensing using wavelengths outside the range included in
flag during conditions when the reflectance data represent a mixture NDVI and WDRVI (Grace et al., 2007; Nakaji et al., 2008).
between green forest canopy and snow. Another issue is related to NDVI and WDRVI may nevertheless be used as a large-scale
poor illumination conditions at high latitude sites in wintertime complement to ground observations of climate change impact on
caused by high solar zenith angles. While the MODIS quality flag data spring phenology, with the strength of the analysis being that satellite
identifies measurements obtained at solar zenith angles above 86°, it data required for calculating NDVI and WDRVI is now available for a
may be necessary to use a more conservative threshold, at least for fairly long period of time. In this study, we showed that the site-
sites located in areas with highly variable topography where specific variation in vegetation indices within different eight-day
shadowing effects are accentuated. periods was related to the temperature dependent timing of
The annual changes in needle biomass are related to the gradual greening-up (including both melting of snow and increase in green
increase by bud burst and shoot elongation during spring and biomass). The relation between phenological indicators and eight-day
summer, and gradual decrease during autumn caused by shedding of periods differed between geographical regions, being most pro-
older needles. The annual turnover decreases with increasing nounced for time periods with high interannual variability. The site-
latitude, from 40% in the South to 20% in the North for Scots pine specific annual curves can be expected to move from relatively stable
and from 20% to 10% for Norway spruce (Lagergren et al., 2006). The seasonal oscillations to more irregular patterns with lower annual
green amplitude did, however, not replicate this pattern, as it was amplitude in response to a progressively warmer climate. Over time
most pronounced in the North where the estimated amplitude was this will induce a shift in the correlation pattern within regions,
influenced by snow melting occurring at the same time as budburst comparable to the latitudinal differences than can be observed today.
(Delbart et al., 2005). Scots pine stands indicated a higher influence The regional analysis requires data from several monitoring sites, as
from snow than Norway spruce stands, and this is most likely related the site-specific analysis is sensitive to uncertainties associated with
to the differences in tree needle densities that affected the satellite monitoring data and curve fitting procedures. Site-specific
proportion of ground visible from the sky. The difference between analysis may also suffer from interference from tree growth, as an
tree species in annual needle biomass turnover was however not annual increase in biomass will cause a subsequent increase in
detected in the South and Mid part, and evergreen understorey vegetation indices, until the saturation level is reached, and this may
vegetation (e.g. mosses, lichens and lingonberry) in stands with falsely be interpreted as an earlier greening. Forest management
lower tree crown density can have caused a dampening of the green practices, such as thinning, may cause the opposite effect unless
amplitude (Wang et al., 2004), whereas NDVI saturation effects can cancelled out by compensatory growth of understorey vegetation
occur in dense forests (Eriksson, 2007). In addition, the phenology of (Vesala et al., 2005).
deciduous understorey vegetation (e.g. blueberry) may affect the The choice of phenological models may influence the interpreta-
seasonal variability observed by satellite monitoring of sparse tion of the satellite data, and temperature driven models were used for
coniferous forest. estimating effects of between-year variations in onset of vegetation
Difficulties in monitoring of coniferous forests call for the period and timing of budburst in this study. The photosynthetic
development of new remote sensing techniques. Although the recovery in spring is temperature dependent (Repo et al., 2006), and a
coniferous trees remain green during the dormancy season, the running five-day average temperature can be used to predict the start
photosynthetic efficiency is down-regulated as a response to low of photosynthesis, though alternating periods with temperatures
demands of carbohydrates (Ensminger et al., 2004). Seasonal changes above and below 0 °C can turn the photosynthetic activity on and off
in photosynthetic efficiency are not visible to the eye, but physiolog- several times (Suni et al., 2003). Different plant species shows a similar
ical changes associated with cold hardening such as clumping of response to annual variations in temperature (Linkosalo et al., 2009;
chloroplasts, modification of cell membranes, cell dehydration and Scheifinger et al., 2002; Sparks & Menzel, 2002), and thermal-time
forming of intracellular ice (Zwiazek et al., 2001) change the light models based on accumulation of degree days after a fixed date has
absorbance and reflectance. These changes could possibly be detected been shown to outperform more complicated phenological models
A.M. Jönsson et al. / Remote Sensing of Environment 114 (2010) 2719–2730 2729

(Linkosalo, 2000; Linkosalo et al., 2008). More sophisticated process- Hänninen, H., & Kramer, K. (2007). A framework for modelling the annual cycle of trees
in boreal and temperate regions. Silva Fennica, 41, 167−205.
oriented approaches may, however, be required to model climate Haussmann, T., & Lorenz, M. (2004). Manual on methods and criteria for harmonized
change impact on tree phenology (Chuine, 2000; Hänninen & Kramer, sampling, assessment, monitoring and analysis of the effects of air pollution on
2007), though the regulation by ambient temperature and light forests Part I Mandate of ICP Forests and Programme Implementation. In. http://
www.icp-forests.org/Manual.htm: United Nations Economic Commission for
conditions is not yet fully understood (Linkosalo et al., 2006). Europe Convention on Long-Range Transboundary Air Pollution International
Co-operative Programme on Assessment and Monitoring of Air Pollution Effects
5. Conclusions on Forests.
Haylock, M. R., Hofstra, N., Tank, A., Klok, E. J., Jones, P. D., & New, M. (2008). A European
daily high-resolution gridded data set of surface temperature and precipitation for
Vegetation index threshold values, commonly used for monitoring 1950–2006. Journal of Geophysical Research, [Atmospheres], 113, D20.
of spring phenology, were not applicable to Swedish coniferous Hird, J. N., & McDermid, G. J. (2009). Noise reduction of NDVI time series: An empirical
comparison of selected techniques. Remote Sensing of Environment, 113, 248−258.
forests. Evergreen trees do not have a sharp increase in greenness
IPCC (Ed.) (2007). Climate change 2007: Impacts, adaptation and vulnerability.
during spring, and the seasonal changes in vegetation indices were Contribution of Working Group II to the Fourth Assessment Report of the
more related to snow dynamics (approximately 90%) than to changes Intergovernmental Panel on Climate Change. Cambridge, UK Cambridge University
Press.
in needle biomass (approximately 10%). The VI-values represent a
Jönsson, P., & Eklundh, L. (2002). Seasonality extraction by function fitting to time-
mixture between green forest canopy and snow covered ground, and series of satellite sensor data. IEEE Transactions on Geoscience and Remote Sensing,
the melting of snow therefore influences the vegetation indices at the 40, 1824−1832.
timing of budburst in boreal forests. A potential for satellite monitoring Jönsson, P., & Eklundh, L. (2004). TIMESAT — A program for analyzing time-series of
satellite sensor data. Computers and Geosciences, 30, 833−845.
of climate change impact on northern coniferous phenology was Jönsson, A. M., Linderson, M. L., Stjernquist, I., Schlyter, P., & Bärring, L. (2004). Climate
indicated by the VI-values for specific eight-day periods being change and the effect of temperature backlashes causing frost damage in Picea abies.
correlated with temperature dependent spring progress (including Global and Planetary Change, 44, 195−207.
Kalberer, S. R., Wisniewski, M., & Arora, R. (2006). Deacclimation and reacclimation of
both melting of snow and increase in green biomass). The regional- cold-hardy plants: Current understanding and emerging concepts. Plant Science,
specific trends were somewhat more pronounced for WDRVI than 171, 3−16.
NDVI, but without any major difference between the two smoothing Karlsen, S. R., Elvebakk, A., Hogda, K. A., & Johansen, B. (2006). Satellite-based mapping
of the growing season and bioclimatic zones in Fennoscandia. Global Ecology and
methods (double logistic and Savitsky–Golay). Biogeography, 15, 416−430.
Karlsen, S. R., Solheim, I., Beck, P. S. A., Hogda, K. A., Wielgolaski, F. E., & Tommervik, H.
Acknowledgements (2007). Variability of the start of the growing season in Fennoscandia, 1982–2002.
International Journal of Biometeorology, 51, 513−524.
Karlsen, S. R., Tolvanen, A., Kubin, E., Poikolainen, J., Hogda, K. A., Johansen, B., et al.
This study was financially supported by the Swedish Research (2008). MODIS-NDVI-based mapping of the length of the growing season in
Council and the Swedish National Space Board. The MODIS data were northern Fennoscandia. International Journal of Applied Earth Observation and
Geoinformation, 10, 253−266.
distributed by the Land Processes Distributed Active Archive Center
Lagergren, F., Grelle, A., Lankreijer, H., Molder, M., & Lindroth, A. (2006). Current carbon
(LP DAAC), located at the U.S. Geological Survey (USGS) Earth balance of the forested area in Sweden and its sensitivity to global change as
Resources Observation and Science (EROS) Center (lpdaac.usgs.gov). simulated by Biome-BGC. Ecosystems, 9, 894−908.
We acknowledge the forest monitoring data from the Swedish Forest Larsen, J. B. (1995). Ecological stability of forests and sustainable silviculture. Forest
Ecology and Management, 73, 85−96.
Agency, the climate dataset from the EU-FP6 project ENSEMBLES Linkosalo, T. (2000). Mutual regularity of spring phenology of some boreal tree species:
(http://www.ensembles-eu.org) and the data providers in the ECA&D Predicting with other species and phenological models. Canadian Journal of Forest
project (http://eca.knmi.nl). Research, 30, 667−673.
Linkosalo, T., Hakkinen, R., & Hanninen, H. (2006). Models of the spring phenology of
boreal and temperate trees: Is there something missing? Tree Physiology, 26,
References 1165−1172.
Linkosalo, T., Hakkinen, R., Terhivuo, J., Tuomenvirta, H., & Hari, P. (2009). The time
Ahti, T., Hämet-Ahti, L., & Jalas, J. (1968). Vegetation zones and their sections in series of flowering and leaf bud burst of boreal trees (1846–2005) support the
northwestern Europe. Annales Botanici Fennici, 5, 169−211. direct temperature observations of climatic warming. Agricultural and Forest
Beaubien, E. G., & Hall-Beyer, M. (2003). Plant phenology in western Canada: Trends Meteorology, 149, 453−461.
and links to the view fromspace. Environmental Monitoring and Assessment, 88, Linkosalo, T., Lappalainen, H. K., & Hari, P. (2008). A comparison of phenological models
419−429. of leaf bud burst and flowering of boreal trees using independent observations. Tree
Beuker, E. (1994). Adaptation to climatic changes of the timing of bud burst in populations Physiology, 28, 1873−1882.
of Pinus sylvestris L. and Picea abies (L.) Karst. Tree Physiology, 14, 961−970. LMV (2002). Produktspecifikation av Svenska CORINE marktäckedata. SCMD-0001.:
Cannell, M. G. R., & Smith, R. I. (1986). Climate warming, spring budburst and frost Lantmäteriet 8 pp.
damage on trees. Journal of Applied Ecology, 23, 177−191. Loman, J. -O. (Ed.). (2008). Statistical yearbook of forestry 2008. Jönköping: Swedish
Chuine, I. (2000). A unified model for budburst of trees. Journal of Theoretical Biology, Forest Agency.
207, 337−347. Menzel, A., Sparks, T. H., Estrella, N., Koch, E., Aasa, A., Ahas, R., et al. (2006). European
Cleland, E. E., Chuine, I., Menzel, A., Mooney, H. A., & Schwartz, M. D. (2007). Shifting phenological response to climate change matches the warming pattern. Global
plant phenology in response to global change. Trends in Ecology & Evolution, 22, Change Biology, 12, 1969−1976.
357−365. Morin, X., Lechowicz, M. J., Augspurger, C., O' Keefe, J., Viner, D., & Chuine, I. (2009). Leaf
Delbart, N., Kergoat, L., Le Toan, T., Lhermitte, J., & Picard, G. (2005). Determination of phenology in 22 North American tree species during the 21st century. Global
phenological dates in boreal regions using normalized difference water index. Change Biology, 15, 961−975.
Remote Sensing of Environment, 97, 26−38. Myneni, R. B., Keeling, C. D., Tucker, C. J., Asrar, G., & Nemani, R. R. (1997). Increased
Eklundh, L., Johansson, T., & Solberg, S. (2009). Mapping insect defoliation in Scots pine plant growth in the northern high latitudes from 1981 to 1991. Nature, 386,
with MODIS time-series data. Remote Sensing of Environment, 113, 1566−1573. 698−702.
Ensminger, I., Sveshnikov, D., Campbell, D. A., Funk, C., Jansson, S., Lloyd, J., et al. (2004). Nakaji, T., Ide, R., Takagi, K., Kosugi, Y., Ohkubo, S., Nasahara, K. N., et al. (2008). Utility of
Intermittent low temperatures constrain spring recovery of photosynthesis in spectral vegetation indices for estimation of light conversion efficiency in
boreal Scots pine forests. Global Change Biology, 10, 995−1008. coniferous forests in Japan. Agricultural and Forest Meteorology, 148, 776−787.
Eriksson, H. (2007). Leaf Area Index of Scandinaviona Forests Methods using in situ and Olofsson, P., Lagergren, F., Lindroth, A., Lindstrom, J., Klemedtsson, L., Kutsch, W., et al.
remotely sensed data. Lund University, Dissertation, 2007. (2008). Towards operational remote sensing of forest carbon balance across
Goetz, S. J., Bunn, A. G., Fiske, G. J., & Houghton, R. A. (2005). Satellite-observed Northern Europe. Biogeosciences, 5, 817−832.
photosynthetic trends across boreal North America associated with climate and fire Parmesan, C. (2006). Ecological and evolutionary responses to recent climate change.
disturbance. Proceedings of the National Academy of Sciences of the United States of Annual Review of Ecology Evolution and Systematics, 37, 637−669.
America, 102, 13521−13525. Persson, G., Bärring, L., Kjellström, E., Strandberg, G., & Rummukainen, M. (2007). Climate
Grace, J., Nichol, C., Disney, M., Lewis, P., Quaife, T., & Bowyer, P. (2007). Can we indices for vulnerability assessments. SMHI Reports Meteorology Climatology (pp. 64).
measure terrestrial photosynthesis from space directly, using spectral reflectance Norrköping, Sweden: SMHI.
and fluorescence? Global Change Biology, 13, 1484−1497. Repo, T., Leinonen, I., Wang, K. Y., & Hanninen, H. (2006). Relation between
Hannerz, M. (1994). Predicting the risk of frost occurrence after budburst of Norway photosynthetic capacity and cold hardiness in Scots pine. Physiologia Plantarum,
spruce in Sweden. Silva Fennica, 28, 243−249. 126, 224−231.
Hannerz, M. (1999). Evaluation of temperature models for predicting bud burst in Rötzer, T., Grote, R., & Pretzsch, H. (2004). The timing of bud burst and its effect on tree
Norway spruce. Canadian Journal of Forest Research, 29, 9−19. growth. International Journal of Biometeorology, 48, 109−118.
2730 A.M. Jönsson et al. / Remote Sensing of Environment 114 (2010) 2719–2730

Saxe, H., Cannell, M. G. R., Johnsen, B., Ryan, M. G., & Vourlitis, G. (2001). Tree and forest Tagesson, T., Eklundh, L., & Lindroth, A. (2009). Applicability of leaf area index products
functioning in response to global warming. The New Phytologist, 149, 369−399. for boreal regions of Sweden. International Journal of Remote Sensing, 30,
Scheifinger, H., Menzel, A., Koch, E., Peter, C., & Ahas, R. (2002). Atmospheric 5619−5632.
mechanisms governing the spatial and temporal variability of phenological phases Tucker, C. J. (1979). Red and photographic infrared linear combinations for monitoring
in central Europe. International Journal of Climatology, 22, 1739−1755. vegetation. Remote Sensing of Environment, 8, 127−150.
Schwartz, M. D. (1999). Advancing to full bloom: Planning phenological research for the Vesala, T., Suni, T., Rannik, U., Keronen, P., Markkanen, T., Sevanto, S., et al. (2005). Effect
21st century. International Journal of Biometeorology, 42, 113−118. of thinning on surface fluxes in a boreal forest. Global Biogeochemical Cycles, 19.
Schwartz, M. D., Reed, B. C., & White, M. A. (2002). Assessing satellite-derived start-of- Viña, A., & Gitelson, A. A. (2005). New developments in the remote estimation of the
season measures in the conterminous USA. International Journal of Climatology, 22, fraction of absorbed photosynthetically active radiation in crops. Geophysical
1793−1805. Research Letters, 32.
Sokal, R. R., & Rohlf, F. J. (1995). Biometry, the principles and practice of statistics in Wang, Y. J., Woodcock, C. E., Buermann, W., Stenberg, P., Voipio, P., Smolander, H., et al.
biological research. New York: W.H. Freeman and Company. (2004). Evaluation of the MODIS LAI algorithm at a coniferous forest site in Finland.
Sparks, T. H., & Menzel, A. (2002). Observed changes in seasons: An overview. Remote Sensing of Environment, 91, 114−127.
International Journal of Climatology, 22, 1715−1725. White, M. A., de Beurs, K. M., Didan, K., Inouye, D. W., Richardson, A. D., Jensen, O. P.,
Stöckli, R., & Vidale, P. L. (2004). European plant phenology and climate as seen in a 20- et al. (2009). Intercomparison, interpretation, and assessment of spring phenology
year AVHRR land-surface parameter dataset. International Journal of Remote in North America estimated from remote sensing for 1982–2006. Global Change
Sensing, 25, 3303−3330. Biology, 15, 2335−2359.
Suni, T., Berninger, F., Vesala, T., Markkanen, T., Hari, P., Makela, A., et al. (2003). Air Zhang, X. Y., Friedl, M. A., & Schaaf, C. B. (2006). Global vegetation phenology from
temperature triggers the recovery of evergreen boreal forest photosynthesis in Moderate Resolution Imaging Spectroradiometer (MODIS): Evaluation of global
spring. Global Change Biology, 9, 1410−1426. patterns and comparison with in situ measurements. Journal of Geophysical
Suzuki, R., Nomaki, T., & Yasunari, T. (2001). Spatial distribution and its seasonality of Research, Biogeosciences, 111, 14.
satellite-derived vegetation index (NDVI) and climate in Siberia. International Zwiazek, J. J., Renault, S., Croser, C., Hansen, J., & Beck, E. (2001). Biochemical and
Journal of Climatology, 21, 1321−1335. biophysical changes in relation to cold hardiness. In F. J. Bigras & S. J. Colombo
Suzuki, R., Nomaki, T., & Yasunari, T. (2003). West–east contrast of phenology and (Eds.), Conifer cold hardiness (pp. 165−186). the Netherlands: Kluwer Academic
climate in northern Asia revealed using a remotely sensed vegetation index. Publishers.
International Journal of Biometeorology, 47, 126−138.

You might also like