Critical sample size and satellite image selection for the recognition of poppy and pyrethrum crops in North West Tasmania

Determination of the critical level of training data, and investigation of targeting the time of image acquisition to specific crop growth cycles, will increase the efficiency of remote sensing data analysis, for recognition of poppies and pyrethrum. The objective of this project, was to determine whether the amount of training data (critical sample size) contributed significantly to classification, using three methods of analysis of two season's data for, poppy and pyrethrum crops, on the North West Coast of Tasmania and to investigate the timing of image acquisition. Distinction between class types was not an objective of the study. Eight Landsat 5 TM, two SPOT XS and three SPOT XI images were acquired between 02 July 1997 and 28 March 1999. Observations of the spectral response of poppies showed that, Landsat TM bands one, two, three and five in November and January, provided significantly different, peak pixel values for the poppy crop. SPOT XS band two in February also provided peak pixel values for poppies. The spectral response of pyrethrum indicated that an increase in pixel value for the January Landsat TM data in bands five and seven was distinct, as was the peak in SPOT XI band three during December. A principal component analysis, (PCA) was carried out separately on each image. For all Landsat TM imagery, over 98.482% of variance was contained within the first three principal components. Similarly, for all SPOT XI data, over 99.189% of variance was contained within the first three principal components. When SPOT XS data was analysed, the first and second components accounted for over 98.909% data variance. The merged spectral response patterns generated from the automatic internal average relative reflectance (AIARR), normalised difference vegetation index (NDVI) and PCA images by an unsupervised, iterative self-organising data analysis technique (ISODATA) for the poppy and pyrethrum AOIs, provided the input for a supervised classification. As the AIARR, NDVI and PCA data were normally distributed, and the spectral response patterns were parametric, a maximum likelihood parametric decision rule was selected. The amount of training data had a significant effect on the contribution to classification for the three analysis methods, over two season's data for each of the crop types. To achieve a classification accuracy of 90% for poppies, the acquisition of Landsat data in either November or January, required a PCA with 80% of the total poppy area used as training (calibration) data. To achieve a classification accuracy of 96% for poppies, the acquisition of SPOT XI data in either October or December, required a NDVI analysis with 50% of the total poppy area used as training (calibration) data. For pyrethrum, a classification accuracy of 80% was achieved by acquiring imagery in the post harvest and dormancy stage (late February to October), using a PCA method and 90% of the total amount of data available for training. When imagery was acquired in late December or early January, using 40% of the total amount of pyrethrum data for training contributed, on average, to the classification of 87% of the crop, when analysed using the NDVI method. The findings of this research showed that the choice of the training set (quality and quantity) had an influence on the success of a classification approach as well as the choice of image analysis technique. Timely acquisition of imagery was shown to be required to achieve a satisfactory level of contribution to classification from training data of poppies and pyrethrum.