Parbat - Screenshots
- Fig 1: Parbat's basic displays
- Fig 2: 3D feature space plot with class spheres
- Fig 3: Visual Fuzzy Classifier explained
- Fig 4: VFC classification result
- Fig 5: Screen capture Parbat session
- Fig 6: Fine tune VFC classication result in 3D plot
- Fig 7: Compare standard FCM with VFC
- Fig 8: Alpha-shapes
- Fig 9: Texture image
- Fig 10: Unsupervised texture-based segmentation
- Fig 11: Semi-supervised texture-based segmentation based on the LBP texture measure
- Fig 12: Multivariate texture-based segmentation of simulated image
- Fig 13: Multivariate texture-based segmentation of hyperspectral CASI image
- Fig 14: Multi-scale LBP texture measure
- Fig 15: Multi-scale LBP texture measure applied to LiDAR DEM
- Fig 16: Class cluster visualisation in 3D feature space
- Fig 17: Isosurfaces for visualisation of classification uncertainty
- Fig 18: 3D Terrain Viewer
Figure 1. The three basic representation forms implemented in our prototype with; A: the image display showing a colour composite of band 4, 5 and 3 (RGB). It uses a pan window to pan the main image display window. The image display contains a zoom window with a slider to set the zoom factor. A separate information window shows the location and values of a selected pixel.; B: The parallel coordinate plot (PCP) shows the signatures of pixels in all seven bands. Transparent histograms give information on the density of pixel values on each axis. The red line shows a brushed pixel signature with labels containing the DN-values per band. The PCP in this figure is a separate Java tool called Parvis (Hauser, 2000).; C: The 3D feature space plot shows the scatter cloud of pixels in band 4,5 and 3 (XYZ). In this density scatter cloud, 2000 of the most occurring pixel values are shown. The pixel brightness depicts their frequency. The plot can be navigated (zoom, pan and (auto)rotate) with the three mouse buttons. This plot can be linked to the image display in figure 1A.
Figure 2. The 3D feature space plot showing reference class spheres of water, pine, maquis, garrigue, agriculture and urban impervious. Overlap occurs between class maquis and pine, maquis and garrigue, and agriculture and urban impervious. The class sphere of the maquis class is selected. The popup window shows the class label, colour, the number of pixels in the reference area, the mean value in each of the bands and the radius of the class sphere. The red pixel in the 3D plot is a selected pixel. Its value is shown in the upper-right corner of the plot window. Additionally, the number of pixels with the selected value in the image is shown. These image pixels are highlighted in the image display.
Figure 3. The interactive visual fuzzy classification is schematically shown in this figure. The agriculture and urban impervious overlap in the 3D feature space plot (A). The class spheres depict the triangular membership functions, which are used for classification. Membership values for unclassified pixels are calculated based on these membership functions. Figure 3B shows the membership functions of the two of the two overlapping classes projected on one bands (band 4). When selecting one of the spheres, a user is presented with a popup window including a slider. This slider is used to adjust the radius of the sphere, thereby changing the width of the membership function and possibly changing the overlap with another class/membership function.
Figure 4. The classification result after an initial visual fuzzy classification based on the default radii of the class spheres. A: the hard landcover class image display with class labels based on the maximum membership. The colours of the classes are the same as the colours used in the 3D plot. The following classes occur: water, pine, maquis, garrigue, agriculture and urban impervious. The black pixels are unclassified, because these pixels occur outside all of the class spheres in the 3D feature space plot. B: the membership image display for the class maquis shows high membership values for pixels in the Northwest, hilly part of the area (bright values depict high membership values). C: the maximum membership map show for every pixel the maximum membership from the membership vector. High values depict low uncertainty (bright), whereas low values depict high uncertainty (dark). The maquis and water pixels are classified with low uncertainty, whereas maquis, urban and agricultural areas with higher uncertainty values (note: high uncertainty = low membership = dark value). D: the confusion image display shows areas where confusion in classification occurs (bright areas). These are the pixels with values within the overlap zone of two or more membership functions. Urban and agricultural areas overlap in feature space, as can be seen in figures 2 and 3. Pixels in this overlap zone show high confusion values in the image display (note: high uncertainty = high confusion = bright value). In future versions, an interactive legend will be implemented to avoid confusion between uncertainty maps.
Figure 5. A screen capture of the “Parbat” prototype, showing several selection, information and display windows. “Parbat” uses a multiple document interface without a backing window. The main window is the button bar in the upper-left corner, with buttons to access the main visualization and classification functions.
Figure 6. After an initial classification based on the default class sphere radii we go back to the 3D plot to fine-tune the classification result. A: the 3D plot with the adjusted class spheres; we increased the size of the water class sphere, as well as the garrigue and agriculture class spheres. We decreased the radius of the urban impervious class. This configuration of class clusters better reflects the situation in the field. B: selection window to add and remove classes to and from the 3D plot. Transparency for the class spheres can be adjusted in this window. The “classify” button initializes another window, figure C, to select the output of the classification. In this window, the classifier is finally started. Figures D, E and F (hard classes, maximum membership and confusion respectively), show the classification result based on the new class sphere configuration.
Figure 7. Comparison of the visual fuzzy classifier with a standard supervised fuzzy c-means. A: the hard class image display based on the class sphere configuration, as shown in figure 6A. B: the confusion index for this class configuration. Transition zones between maquis and garrigue show high confusion, as well as some urban and agricultural areas. C: the hard classes for the standard fuzzy c-means classifier. It shows that the urban impervious class is over-classified and the agricultural class is under-classified. High confusion values occur in agricultural, urban impervious and garrigue areas.
Figure 8. Class Alpha-shapes in 3D feature space plot
Figure 9. Simulated texture image on the left side with training areas for classification. The right image is an example of a pixel-based classification. The different textures are not identified by this pixel-based approach.
Figure 10. The left image shows the result of an unsupervised split-and-merge segmentation of fig. 9 based on segment mean and variance values only. The right image is the result of an unsupervised split-and-merge segmentation based on the LBP texture measure.
Figure 11. The left image shows the results of a semi-supervised split-and-merge segmentation of fig. 9 based on the LBP and variance texture measures. The right image show uncertainty related to segment class assignment. As expected, high uncertainty values occur on the boundaries of textures.
Figure 12. The example in fig. 9 is extended to a multi-band (multivariate) texture-based image segmentation. It shows that textures at different scales and of different colours can be identified.
Figure 13. The multivariate texture-based segmentation algorithm is applied to a hyperspectral CASI image of the Ainsdale Sands, UK.
Figure 14. The LBP texture measure is extended for multiple scale levels.
Figure 15. The results of a multi-scale LBP texture segmentation of a LiDAR DEM of the Ainsdale Sands, UK.
Figure 16. Visualisation of a class cluster in 3D feature space using different shapes: a. class pixels, b. sphere, c. ellipsoid, d. convex hull, e. alpha-shape, f. isosurface. The isosurface is very fast to compute and render and allows for fast interaction.
Figure. 17 Multiple isosurfaces for three class clusters based on a user-defined uncertainty threshold. The lower the uncertainty threshold the larger the isosurface (left) and spatial extent of the classes (right). Isosurface visualisation is used to explore the results of a soft/fuzzy classifier. Uncertainty in the spatial extent of a class is closely linked to uncertainty in the thematic or spectral extent of a class in feature space.
Figure 18. Parbat's 3D Terrain Viewer