ISPRS Archives XXXVIII-8/W3 Workshop Proceedings: Impact of Climate Change on Agriculture 
327 
algorithms for filling the data gaps and removing data anomalies 
are required. Such a pre processing step is also called as 
construction of time series and various methods have been 
proposed to construct the time series to match the ground process. 
(Colditz et.al.2008; Jonsson and Eklundh, 2002,2004). Popularly 
used techniques include minmax filtering/temporal window 
operation, savitzky golay filtering, fourier series fitting, gaussian 
model fitting and logistic function fitting. We have used local 
maximum fitting devised by Sawada et.al., (2005) which combines 
temporal window operation and fourier series fitting. 
Time 
Figure 1. Effect of LMF Filtering 
3. METHODS 
The various steps involved in deriving the variability map and 
cropping season map are described in the flow diagram of figure 2. 
As explained in section 2.1, the time series of EVI is generated and 
the pre-processing steps are applied. The output of this step 
including raw and smoothed time series for all pixels is input to the 
time series classifier. 
3.1 High level classification 
This paper builds on the time series classification process 
developed by the authors the results of which are yet to be 
published. For the purpose of this paper, it will be assumed that 
crop pixels have been identified. Since, we are interested in the 
cropping practices, only crop pixel-time trajectories are subjected 
to further processing. The main function of this step is to eliminate 
non crop pixels. 
3.2 Time series analysis 
A number of methods have been developed to determine the start 
and end of growing seasons using time series of vegetation indices. 
These methods have employed a variety of different approaches 
including the use of specific NDVI thresholds (Lloyd, 1990; White 
et al., 1997), the largest NDVI increase (Kaduk & Heimann, 1996), 
backward-looking moving averages (Reed et al., 1994), or 
empirical equations (Moulin, et al., 1997). In our work, we have 
determined the inflection points of each crop pixel-time trajectory 
and determined the phenological parameters as defined by Jonnson 
and Eklundh. (2002,2004). The definitions of the phenological 
parameters extracted are 1) Time for the start of the season: 
datestamp at which there is a rise of 20% above the left minima. 2) 
Time for end of the season: datestamp at which the right edge is 
20% above the measured right minima 3) Time for the mid of the 
season: computed as the mean value of the times for which, 
respectively, the left edge has increased to its 80 % level and the 
right edge has decreased by 80 % level; and 4) Seasonal amplitude: 
difference between the maximal value and the base level. These 
parameters are illustrated in figure 3 
Figure 2. Overall Flow Diagram of the Process