study 
peas, 
ccoli, 
t and 
uo um f 
FF 
£ 
  
5 
Barrett, Rachel 
  
2 MATERIALS AND METHODS 
Digital data from SPOT-XS, SPOT-XI and Landsat TM were acquired over the study area during the spring/summer 
growing seasons of 1997/98 and 1998/99. These data were processed, classified and compared to ground-truthed data of 
crop cover as described below. 
2.1 Collection of Ground Based Data. 
Ground-based data were collected at fortnightly intervals during these growing seasons. The 17 crop types selected for 
recording based on commercial significance, were oats (OA), triticale (TR), wheat (WH), barley (BA), onions (ON), 
carrots (CS), peas (PE), beans (BE), potatoes (PO), pumpkin (PU), broccoli (BR), poppies (PP), pyrethrum (PY), 
cauliflower (CA). sweet corn, lucerne and pasture. The collected data assessed and recorded crop type, growth stage and 
ground cover from randomly selected field areas representing approximately 10% of the total area under cultivation. 
Crop boundaries were recorded utilising 1:5000 orthophoto map series. 
2.2 Data Acquisition and Initial Processing. 
Digital data from SPOT-XS, SPOT-XI and Landsat TM systems were acquired, for several reasons. As stated by Reid 
et al (1993), the high incidence of cloud over Tasmania makes the acquisition of cloud-free images via the Landsat TM 
system alone somewhat difficult. These limitations occur elsewhere (Stoney, 1996), (Wiegand' al, 1996). The SPOT 
system increases the ability to acquire cloud-free images by the use of off-nadir viewing. The use of both Landsat TM 
and SPOT increases the frequency of target overpass compared with the use of one system only. 
Subsequent to image acquisition, and using ERDAS Imagine Version (8.3) the images were resampled to produce 
uniform 20 metre square “pixels”, and to accurately align the images to one another. The images were then “stacked” to 
make one image. This made it possible to analyse all the information for the growing season at one time. Crop 
boundaries, which had been manually digitised from 1:5000 orthophoto map series were converted to digital form and 
then incorporated as a mask, so that information outside of the crop boundaries (areas of interest, AOI) were masked out 
to black. 
For the 1997/98 growing season, six images were acquired, between 02 July 1997 and 18 April 1998. The sequence was 
as follows; TM 02 July 1997, TM 06 October 1997, TM 23 November 1997, TM 10 January 1998, XS 23 February 
1998 and XS 18 April 1998, providing a total of 30 measurements in each “pixel” of the AOI. For the 1998/99 growing 
season, seven images were acquired, between 21 July 1998 and 03 March 1999. The sequence was as follows; TM 21 
July 1998, TM 27 September 1998, XS 08 October 1998, XI 31 December 1998, TM 29 January 1999, TM 03 March 
1999 and XI 29 March 1999, providing a total of 35 measurements in each “pixel” of the AOL 
2.3 Image Classification 
The software package chosen was ERDAS Imagine Version (8.3) Figure 3, represents diagrammatically the image 
classification process. The first step in the analysis procedure was a Principal Components Analysis (PCA) executed on 
the “stacked”, masked images incorporating all the season’s image data. This procedure is a standard procedure to 
concentrate the significant information into a smaller number of bands. About a third of the bands were discarded after 
this procedure because they mostly contained noise. 
An unsupervised classification process (maximum likelihood), then noted which pixels were similar and formed classes 
for similar pixels. Five hundred classes were utilised in the 1997/98 season and 400 in the 1998/99 season. Classes were 
used to "classify" the whole-unmasked area. After this classification process, a file was produced which contained the 
details of the average values for each of the classes. All pixels for each paddock could then be represented by a class, to 
which the pixel belonged. 
The data was then prepared for entry into a custom computer program (“PaddockId”, compiled in C Program language). 
The PaddockId program had three input files. The first was the information from an unsupervised classification, 
containing the average values of the classes. The second file listed, for each paddock, the number of pixels in each of 
these classes. The program also required a training dataset. About one third of the field areas for each crop type 
recorded in the ground-truthing was used as "training data", which was used to “learn” the characteristics of the crops 
and their variation with time. These were called the "known field areas". The remainder, were the "unknowns", and 
analysed by the procedure and used to score the accuracy of the process. The third file listed the class types expected, 
the identification numbers of the "known" field areas and their crop types, the paddock identification numbers of any 
field areas affected by cloud, and some probability limits used by the program to decide when field areas could be 
considered to be similar. 
The program used information extracted from the “stacked”, masked image. Several runs of this program were 
conducted, using all of the bands of information for one year. The amalgamation of the results from the various 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 135