Satellite Applications for Detecting Vegetation Phenology - Satellite-based Applications on Climate Change - page 258

wide variety of cover types that occur in the continental United States. Reed et al.
( 1994 ) developed an automated, quantitative approach to derive phenological
measures from multitemporal Advanced Very High Resolution Radiometer
(AVHRR) NDVI observations. To identify the onset of the growing seasons, an
auto-regressive moving average of previous smoothed nine NDVI biweekly com-
posite values was compared to the smoothed NDVI value. Selecting the moving
average time interval (the number of NDVI composite periods used to calculate the
moving average) is a critical issue; a large time interval may miss natural vegetation
changes, while a small interval may result in extremely noisy NDVI curves. White
et al. ( 1997 ) provided a methodology which determined the start and the end of
growing season by the threshold of the normalized NDVI ratio. Instead of original
NDVI values, they normalized the NDVI to 0-1 by its maximum and minimum
value. A new smoothed NDVI ratio curve was developed based on the method of
Reed et al. ( 1994 ), and then the NDVI ratio threshold of 0.5 was used to identify
growing season length. Moulin et al. ( 1997 ) used time derivative of NDVI to detect
three transition dates of vegetation cycle: beginning, maximum, and end. The time
derivative before beginning date should be zero and after beginning date should be
positive; the end date was calculated similarly to the beginning date. The algorithm
is sensitive of the weight of the derivative term. If the weight is too large, the
detection may be confused by short-term signal variations due to residual noise
(e.g., soil color, directional effects). If the weight is too small, the algorithm may
fail for pixels, which remain partly green during the year. Duchemin et al. ( 1999 )
revealed that the temporal variation of NDVI during budburst and senescence was
nearly linear. A line segment model was used to fit the effect of budburst and
senescence. The method was sensitive to a change in the rate of NDVI variation,
resulting, for instance, from a spring frost during budburst or from a severe drought
in summer accelerating the senescence. Zhang et al. ( 2003 ) identified phenological
transition dates based on the curvature-change rate of a logistic model for time
series of MODIS vegetation indices. This method has been applied in many
researches (Ahl et al. 2006 ; Peckham et al. 2008 ; Zhang et al. 2004 ) because it is
able to handle multiple growth cycles and is not tied to a specific calendar period
(e.g., January to December). The challenge for this method is to identify a single
sustained increase (growth) and decrease (senescence) period before the MODIS
measurements could be fit to the logistic model.
Li et al. ( 2010b ) modified Zhang et al. ( 2003 )'s approach by using the Fourier
series to decompose the periodic vegetation indices instead of the moving average
window Eq. 17.3 .

y ¼ a þ b cos ðωxÞþc sin ðωxÞ (17.3)

where a , b , and c are fitting parameters; ω is the angular frequency which is equal to

2 π = T ; and T is period time of the growth cycle. For instance, the T for the single

growth cycle is the number of data samples for 1 year. For multiple growth cycle,
T equals to the number of data samples divided by the number of cycles for 1 year.
The other parameters are solved by least square fitting. Local maximum and
minimum points of the simulated data divided original data into a series of
sustained increasing and decreasing trends (Fig. 17.1 ). Using Fourier series to