Page 3 of 3 Previous It is noted that some training sites were selected to cover two classes on purpose. By doing this, a training site may contain more than one class and create the fuzzy situation for training data. The detailed training statistics are presented in table 2.  The DN values of the training sites were plotted on the histogram and their relative frequency calculated. After that, using the equation (2), the membership grade of the classes were determined for that selected site. When the number of training sites for each class were more than one, their membership grades were averaged. Using these membership grade values the partition matrix was created which was shown in table 3.  The next stage of fuzzy supervised classification was to create the fuzzy signatures. Here the main steps were specify the name of the file that defined the signatures, the number of bands to use in signature development, the names of the bands and the names of the signatures to be created.  Fig. 3 Fuzz signature of the classes. 5.4 Supervised classification The last stage of the fuzzy supervised classification was to classify the ASTER image with the fuzzy signature group file. In this case, conventional maximum likelihood prior probability classier was used only replacing the conventional mean and covariance matrix by fuzzy mean and covariance matrix. The equations (5) and (6) described this procedure. The final classified ASTER image is shown in figure 4. 6. Results and Discussion The classified image revealed that all land cover classes were mixed in nature, except non-forest (bare ground) class. The pine forest class was heavily scattered in nature, which contradicted the field survey data. On the other hand, other classes intermixing were acceptable, in context of their distribution characteristics. The area and percentage of the land cover classes were shown in table 3.   Fig.4 Fuzzy classified ASTER image of the study area. 7. Accuracy Analysis The objective of the accuracy assessment is to estimate the degree of correct classification of the final map. The procedure involves a comparison between two sources of information: the remote sensor derived classified map and a reference map. The classified map is considered accurate if it corresponds closely to the reference map (Congalton et al, 1983). Source of errors in classification may include misidentification of land pattern (commission errors) and excessive generalization (omission errors). In this study, accuracy assessment involved four procedures. They are the determination of the spatial sampling size, reference data collection, creation of error matrix and accuracy analysis. 7.1 Spatial Sample Size The sample size refers to the number of pixels to be visited (or referenced) on the ground and used to assess the accuracy of individuals land cover categories on the classified map. Various statistical methods are applied to determine the sample size to validate accuracy assessment. Congalton (1991) suggests that “ a good rule of thumb is to collect a minimum of 50 samples for each land cover category”. However, to provide an adequate redundancy for the accuracy assessment, the number of sample points were 343, which was more than 250 points for 5 classes. 7.2 Reference Data Collection Reference data was collected from the aerial photographs and the field survey. These aerial photographs are at scale of 1:25,000 and have been visually interpreted. Some sample reference data for the accuracy assessment were collected from field survey with the help of detailed topographic maps and hand held Global Positioning System (GPS). The reference sample point locations were identified on the ground and then the surrounding land cover types were noted down. 7.3 Accuracy Assessment Table 4 shows the error matrix for ASTER classified land cover map which was generated from the sampling points for the classified satellite data and reference data. Within an error matrix, omission errors are caused by the misclassification. These measurements are computed by dividing the total number of correct pixels in a category into the total number of that category as derived from the reference data (table 4, row values). On the other hand, commission errors, are caused by the analyst, and are computed by dividing the total of correct pixels on a category into the total number of that category obtained from classified map (table 4, column values).  7. Conclusion The fuzzy supervised classification results show that ASTER data are suitable for forest cover mapping. Even though only three visible bands were used for fuzzy classification, the 15m image classification results revealed the clear distribution of forest species. This study showed that fuzzy classification works more efficiently on fine resolution images than other coarse resolution images. A pixel is no longer considered as an indecomposable unit in image analysis. Information about pixel component cover classes becomes available, and a pixel’s partial membership enables more accurate statistical parameters. The described partial membership value generation method is a very convenient method to determine membership values from heterogeneous training sites. In information processing, sometimes efforts are mainly made on algorithms and neglect the knowledge representation. As a result, sophisticated algorithms can produce inferior results for poor knowledge representation. This histogram method of membership value generation will provide a better way to represent a membership of a pixel in fuzzy supervised classification. Acknowledgement The authors would like to thank Mr. Ross Dodds, of the Satellite Remote Sensing Service (SRSS) Center at the Department of Land Administration of Western Australia (DOLA), for providing newly acquired ASTER data of Albany region for this study. References
Page 3 of 3
Previous |