382 
components. When maximum values of greeness for corn 
and soybean were plotted against time it was found 
that a simple linear classifier could separate the 
two. 
This work carried out in the United States shows 
clear advantages in using multitemporal Landsat data 
in crop classification, though the range of crops 
considered has been limited with a concentration of 
work on corn - soybean separation. The range of crops 
grown in an area, the field pattern, and most 
importantly the timing of the image within the 
agricultural calendar can be identified as the three 
key factors limiting the sucess of crop 
classification from Landsat. 
In the West European agricultural situation, crops 
that are phenologically very similar such as winter 
wheat, winter barley and spring cereals are often 
produced in adjacent blocks. When this is combined 
with the production of oilseeds, root and vegetable 
crops, over very small geographical regions then the 
picture presented to the scanner is one of great 
spectral complexity. However, if Landsat MSS image 
aquisition can be matched to key points of change in 
the crop calendar and combined as multitemporal data 
sets, then the differences in the temporal-spectral 
profiles of even phenologically similar crops may be 
sufficient for accurate crop identification. 
The sucessful application of the temporal dimension 
as an aid to crop classification in the United States 
suggests that a similar approach may improve the crop 
classification accuracy in the more complex West 
European agricultural situation. To investigate this, 
two main objectives were defined, firlstly to 
identify the optimum single * date for crop 
classification, and secondly to examine changes in 
crop classification accuracy through the use of 
single date and multitemporal data with reference to 
the United Kingdom agricultural situation. 
2. EXPERIMENTAL METHODS 
A test site was established in an area of diverse 
cereal, oilseed and vegetable crop production in 
Southern England. The exact location was an area of 
625 square kilometers centred on Framlingham, 
Suffolk. Some 2000 Ha of ground data was available in 
the form of farm records showing field boundaries and 
crop type. 
A search through the National tape archive for 
Landsat data from 1972 - 1984 was carried out. Table 
2.1 shows all available images of reasonable quality 
for the Framlingham area. 
Table 2.1 Image availability 
Year Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 
1974 
1975 
*** 
1976 
*** 
*** 
1977 
*** 
1978 
•kkk kkk 
1979 
*** 
*** *** *** 
1980 *** *** 
*** 
kkk kkk 
*** 
1981 
*** 
*** 
1982 *** 
•kick 
kkk 
kkk kkk 
This illustrates the poor temporal coverage of 
Landsat data. Only the 1979 / 1980 growing season 
offered any spread of images through one growing 
season and even this is far from ideal. A notable gap 
exists in the late June, early July period which 
corresponds to the scenescence period of early 
ripening crops and as such could be vital in 
separating closely matched crops like winter wheat 
and winter barley. The useable images were: 2nd 
November 1979 and 18th February, 12th April, 18th 
May, 4th June and 16th August, 1980. All are path 
217, row 23 except the image from 4th June 1980 which 
is path 216, row 24. 
A fundamental pre requisite for the use of 
multitemporal data is the ability to register the 
imagery to a common projection, either image to 
image, or image to map. This is neccessary for pixel 
by pixel comparisons from one date to another. The 
accuracy of any results derived from multitemporal 
data sets are closely related to the accuracy of 
registration of the original scenes, (Anderson, 
1985), so much effort must be put into ensuring a 
high degree of precision when transforming the 
original imagery. 
Correction of each scene was carried out using the 
interactive geometric correction package on a GEMS 
image processing system based on a PRIME 2250. This 
involves the matching of image points with known 
ground locations as found on 1:50,000 scale O.S. maps 
and using the corresponding pairs of coordinates to 
generate the coordinate transform equations. 
Approximatly 20 control points in the vicinity of the 
study area were used for each scene. 
The transform equations for each of the 6 scenes, all 
linear, were applied to the whole scenes with average 
ground location errors being less than 100m for each 
image. During the transformation each image was 
resampled to give a 50m pixel, interpolation was 
carried out using the Bilinear interpolation 
algorithm. 
A 1000 pixel by 600 line geometrically correct 
subscene running from minimum easting, 603000, 
minimum northing, 254000, to maximum easting, 653000, 
maximum northing, 284000, was produced from each date 
for the Framlingham test site. Cloud and haze in the 
February and August images caused severe degradation 
to the quality of these data. The June data suffered 
from random noise which proved difficult to remove. 
From examination of the band 4,5,&7 false colour 
composites for each image a 512 pixel by 512 line 
extract was defined to coincide with maximum ground 
data availability and good image quality. 
Ground data for each crop was matched to the imagery 
by converting the easting and northing coordinates 
for each field from the mapped ground data into X and 
Y image coordinates. The classification proceedures 
in the available software were then used to locate 
each field and draw in a training area. The training 
areas drawn for each field matched for all images. 
This was possible for the November, February, April 
and May scenes but the cloud and haze on the August 
scene and the bad data of the June scene meant that 
seperate training sets had to be established for 
these two dates. Note however that no new areas were 
introduced but that unsuitable areas were removed. In 
all, 10 training sets were established, one for each 
of the ten cover classes: winter wheat, winter 
barley, spring barley, oilseed rape, sugar beet, 
peas, beans, and coniferous and deciduous woodland, a 
total area of 2000 Ha. 
Where 24 bands of multitemporal Landsat data are 
available for crop classification, as in this study, 
the identification of the key bands for crop 
discrimination becomes vital. This process was 
accomplished through the use of a decision tree 
applied to coincident spectral plots. 
Mean pixel values and standard deviations for each of 
the 10 cover types were extracted by using the 
previously defined training areas. These data were 
collected for all bands of all images and plotted as 
coincident spectral plots, after Lindenlaub and 
Davis, 1978.