very difficult. Based on annual variations of VI time series, phenological parameters such as dates of the start of season (SOS) and end of season (EOS) and
consequently the growing season length can be estimated. Additional parameters
may also include the dates when vegetation activities reach a stably high level in
the summer (maturity) and a subsequent transitional autumn phase when foliage
starts to decline (senescence) (Zhang et al. 2001).
Various methods have been adopted to estimate the threshold points or phenologically important dates using multi-temporal VI imagery (Lloyd 1990; Reed
et al. 1994; White et al. 1997, 1999, 2002; Zhang et al. 2001, 2003). An early
attempt using fixed threshold values was made to mark the SOS in selected ecosystems (Lloyd 1990). Lloyd (1990) employed a constant value of 0.099 for
AVHRR NDVI time series as the threshold to mark the start and end of the
growing season. This threshold selection was based on tests in selected regions and
the general assumption was that vegetation growth is unlikely to be occurring if
the NDVI value is below the specified threshold (Henricksen and Durkin 1986;
Justice et al. 1986). Given that the same criterion is applied to all land cover types,
geographic variations of LSP may be highlighted using this approach. However,
whether a unified threshold corresponds to the vegetation greenness onset for a
large region is difficult to verify. The assumption of land cover homogeneity puts a
fundamental limit to simple threshold-based phenology for representing spatially
variant biophysical reality in vegetated landscapes. Depending on the vegetation
types and background reflectance properties, the NDVI threshold marking SOS
may be different across locations. Hence, more accurate determination of phenological transition points and growing season length requires land cover/pixelspecific extraction methods that account for the heterogeneous nature of LSP.
More accurate and also more computationally intensive approaches were later
developed in the same era marked with prevalent vegetation studies using AVHRR
NDVI, especially driven by an increased need to monitor LSP for continental and
global-scale coverage (Reed et al. 1994; White et al. 1997). Reed et al. (1994)
derived phenological metrics using primarily a delayed moving average (DMA)
method and compared the results for agricultural crops, grasslands, and forests.
Biweekly composited NDVI time series were interpolated to remove gaps between
adjacent data points, and cloud contamination was further removed using a running median line-smoother. Moving averages of previous n observations (sum of
all values divided by the time interval) were calculated for all pixels. Through
repetitive testing, a nine-value composite period was selected and new DMA time
series were generated. A new time series represents a predicted trend based on
NDVI values in the past. The time when an observed value exceeds the predicated
trend (smoothed NDVI curve crosses the DMA curve) was designated as the onset
of greenness or start of the growing season (SOS). This crossing indicates an
abrupt change of vegetation activities that have surpassed the preexisting tendency. A chronologically reversed DMA procedure was performed in like manner
to derive a metric for the end of the growing season.
White et al. (1997) developed a NDVI ratio-based approach for deriving phenological markers. Instead of using biweekly composite NDVI data, the raw data
106
J. M. Hanes et al.
consequently the growing season length can be estimated. Additional parameters
may also include the dates when vegetation activities reach a stably high level in
the summer (maturity) and a subsequent transitional autumn phase when foliage
starts to decline (senescence) (Zhang et al. 2001).
Various methods have been adopted to estimate the threshold points or phenologically important dates using multi-temporal VI imagery (Lloyd 1990; Reed
et al. 1994; White et al. 1997, 1999, 2002; Zhang et al. 2001, 2003). An early
attempt using fixed threshold values was made to mark the SOS in selected ecosystems (Lloyd 1990). Lloyd (1990) employed a constant value of 0.099 for
AVHRR NDVI time series as the threshold to mark the start and end of the
growing season. This threshold selection was based on tests in selected regions and
the general assumption was that vegetation growth is unlikely to be occurring if
the NDVI value is below the specified threshold (Henricksen and Durkin 1986;
Justice et al. 1986). Given that the same criterion is applied to all land cover types,
geographic variations of LSP may be highlighted using this approach. However,
whether a unified threshold corresponds to the vegetation greenness onset for a
large region is difficult to verify. The assumption of land cover homogeneity puts a
fundamental limit to simple threshold-based phenology for representing spatially
variant biophysical reality in vegetated landscapes. Depending on the vegetation
types and background reflectance properties, the NDVI threshold marking SOS
may be different across locations. Hence, more accurate determination of phenological transition points and growing season length requires land cover/pixelspecific extraction methods that account for the heterogeneous nature of LSP.
More accurate and also more computationally intensive approaches were later
developed in the same era marked with prevalent vegetation studies using AVHRR
NDVI, especially driven by an increased need to monitor LSP for continental and
global-scale coverage (Reed et al. 1994; White et al. 1997). Reed et al. (1994)
derived phenological metrics using primarily a delayed moving average (DMA)
method and compared the results for agricultural crops, grasslands, and forests.
Biweekly composited NDVI time series were interpolated to remove gaps between
adjacent data points, and cloud contamination was further removed using a running median line-smoother. Moving averages of previous n observations (sum of
all values divided by the time interval) were calculated for all pixels. Through
repetitive testing, a nine-value composite period was selected and new DMA time
series were generated. A new time series represents a predicted trend based on
NDVI values in the past. The time when an observed value exceeds the predicated
trend (smoothed NDVI curve crosses the DMA curve) was designated as the onset
of greenness or start of the growing season (SOS). This crossing indicates an
abrupt change of vegetation activities that have surpassed the preexisting tendency. A chronologically reversed DMA procedure was performed in like manner
to derive a metric for the end of the growing season.
White et al. (1997) developed a NDVI ratio-based approach for deriving phenological markers. Instead of using biweekly composite NDVI data, the raw data
106
J. M. Hanes et al.