Examples of satellite image segmentation using FCM-HyDyGAIn this sub section, FCM-HyDyGA cooperative method is used to segment two differentsatellite images : a medium spatial resolution image (30 m) and a high spatialresolution image (4 m). These two images are captured by Landsat ETM+ andIKONOS respectively. It should be noted here that both images consist of multibands,but only three spectral bands are used.
In Addition, IKONOS and Landsatsatellite images are pan-sharpened with high to very high resolution panchromaticbands. IKONOS image consists of many bands one of them is the panchromaticband. To improve the spatial resolution of the visible to near infrared (VNIR), thesebands are fused “pan-sharpened” with panchromatic band (1 meter) (Figure 9a). Inthe case of Landsat the panchromatic the fusion process increases the resolutionto 15 meters (Figure 9b). HyDyGA is initiated with 30 iterations, but this can beincreased without any dropping with the efficiency linked to the speed of convergence for the new method. Being a semi-supervised method FCM requires that thenumber of clusters be provided a priori. To overcome this obstacle, the new methodruns HyDyGA first.
This means that the optimal number of clusters and their valuescan be obtained before running FCM (optimal global solution). Segmentation of theimages by FCM-HyDyGA resulted in different classes for each image such that inthe case of IKONOS image only 4 (Figure 9c) classes are obtained while in the caseof Landsat image only 6 classes are obtained (Figure 9d). Three important classesin the segmented Landsat image are selected for evaluation 1 – urban settlements(black color), 2- bare soil (dark blue), 3- agriculture (dark green), while all the fourclasses are evaluated in the case of IKONOS image 1- Vegetation 1 (dark blue), 2-Urban settlements (red), 3- Soil (orange), 4- Shadow (Light blue). The results ofthe segmentation method FCM-HyDyGA are evaluated using the confusion matrix47 which consists of actual (field samples) and predicted (segmentation results)values. Table 5 and 6 show the evaluation of both Landsat and IKONOS images.Normally, the distribution of the field samples play an important role in the accuracy reliability. The distribution of these samples in this experiment ranges betweenuniform to random depending on the accessibility of the area of studies.
FCM-HyDyGA confusion matrix for Landsat 7 ETM+Using the confusion tables to evaluate the segmentation results of both Landsatand IKONOS by FCM-HyDyGA, one can determine that the accuracies of bothsegmented images to be 97% and 90% respectively. However, one can argue thatthe number of samples are not equal and at the same time not all the classes ofLandsat images are included. But, considering all these factors we still believe thatobtaining an accuracy equal 90% for IKONOS and higher for Landsat is appropriateto prove the high efficiency and reliability of FCM-HyDyGA in the providing aglobal optimal solution.FCM-HyDyGA matrix for IKONOS