a- 
on 
re is a 
asy to 
nporal 
d here 
lataset 
to the 
nonics 
. Crop 
acy of 
tations 
994), 
and 
gical 
been 
sonal 
ourier 
f the 
es in 
n two 
ourier 
with 
Loess 
based 
(Sun, 
»cated 
of 12 
ouble- 
vheat, 
winter 
| crop 
yosites 
lataset 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
obtained from TM images, Chinese general agricultural 
regionalization data, Chinese agricultural cultivated system 
regionalization map and ground based agricultural 
phonological observation data. 
A time series of 10 days composites NDVI derived from 8km 
AVHRR using the Maximum Value Composite (MVC) 
technique were used for analysis crop rotations in study areas. 
The data were obtained from the NASA Pathfinder AVHRR 
Land dataset. The 36 10 days composites NDVI which cover 
the two regions were created by averaging the same period data 
in 1999 and 2000 to reduce climatic influence, and 
corresponding average data of 1997 and 1998 were used to 
instead of missing periods from November to December both 
in 1999 and 2000. The images which are originally in the 
Goode's Interrupted Homolosine Projection were reprojected 
to Albers Equal Area Map Projection before being used in the 
research. 
Cropland grid data with a spatial resolution of 8km were used 
as a mask to extract NDVI time series for cropland in study 
areas. The data were aggregated based on the maximum 
percentage of land cover in each cell from 1km NLCD- 
1999/2000 (Liu, 2003), which were derived from NLCD-1996 
(Liu, 1996) through updating with Landsat TM. 
The ground based agricultural monitoring data derived from 96 
local stations in year 2000 (Fig.l) including the crop 
phonological calendar and main crop types planted in the local 
area were used for evaluation. The Chinese cultivated system 
regionalization map (Chinese agriculture regionalization 
committee, 1991) was used as other kinds of dataset to assess 
the result. 
  
Figure 1. The location of study area in Chinese general 
agricultural regionalization map, the distribution of 
cropland in China at 8km resolution and the 
distribution of 96 local stations. 1. Loess Plateau 
Region (HT) 2. Yellow-Huai-Hai Rivers’ Region 
(HH). 
3. METHODS 
3.1 Discrete Fourier Transformation 
Discrete Fourier transform (DFT) as a signal decomposition 
method can decompose discrete temporal data to the frequency 
235 
domain. The discrete Fourier transform is given by Eq.(1) 
(Moody, 2001): 
Sm LAN (1) 
1 
PTs 
y N 59 
where N = the number of samples in the time series 
k = an index representing the current sample number 
i = an imaginary number 
c = the kth sample value. 
By the DFT a time-dependent periodic phenomenon can be 
decomposed into a series of constituent sine and cosine 
functions and can also be converted to the sum of an additive 
and a series of sinusoidal waves (harmonics, or orders). The 
additive is the arithmetic mean and each wave is defined by a 
unique amplitude and phase angle, where the amplitude value 
is half the height of a wave, and the phase angle (or simply, 
phase) defines the offset between the origin and the peak of the 
wave over the range 0- 2x for the first harmonic, 0- 4x for the 
second harmonic and so on. Successive harmonics are added to 
produce a complex curve, and each component curve, or 
harmonic, accounts for a percentage of the total variance in the 
original time-series data set. The majority of the variance in a 
data set is contained in the first few harmonics. 
In this study the discrete Fourier transform was applied to-the 
36 10 days composites AVHRR time series NDVI dataset on a 
per pixel basis for the whole cropland of study areas. Images of 
the additive, and amplitude and phase angle for each harmonic 
to the eighteenth harmonic were produced on a per-pixel basis 
for each pixel in the NDVI dataset. The amplitude and phase 
were presented in the unit of NDVI and 10 days. Percent 
variance of each harmonics was computed and then the 
additive and the first four harmonics (amplitude and phase) 
were extracted and used for further analysis because about 87% 
of the variance is captured in the additive and the first four 
harmonics. 
3.2 Classification 
Unsupervised classification method was used in this study. 
Image of the additive, and amplitude and phase for the first 
four harmonics (13 bands together) was used as input to the 
iterative ISODATA clustering algorithm, and a convergence 
threshold of 95% and 10 as the maximum number of iterations 
were assigned. Twenty spectral clusters were generated for 
each round and multi-rounds of classification were performed 
for the image to account for mixed clusters. Clusters were 
merged and labeled to four crop rotation classes: single 
cropped for paddy rice and others (non-paddy rice such as 
wheat, maize, soybean, rapeseed etc.), double cropped for 
others/rice and others/others, double cropped rice and triple 
cropped rotations were not considered in this paper because 
none of these situations occurs in these areas. These four crop 
rotations classes were assigned based on the analyst's 
knowledge of agriculture and patterns constructed from the 
summing of amplitudes and phases mean of first four 
harmonics and paddy fields data aggregated from NLCD- 
1999/2000 dataset.