Use of Satellite Imagery for Updating Terrain Information
in Topographic Databases
The Model for Image Segmentation
The major point of interest is to detect linear features from SPOT MSS imagery. Linear features detection can be identified as a process for detection of the region heterogeneity in the image. Image heterogeneity can be measured considering local image coefficient variation. Therefore, the distribution of the reflectance values of homogeneous area has to be modeled and deviation from this model can be considered as heterogeneities.
The point source on the ground is seen on an image as a diffused circular region, due to the properties of the optical involved in imaging. A cross-section of the recorded or imaged intensity distribution of a single point source is shown in figure 1.0 and it can be seen to having a Gaussian-type distribution (Paul, 1987). The actual shape will depend upon the properties of the optical components of the system and the relative brightness of the point source. The distribution function shown in figure 1.0 is called the Point Spread Function (PSF) (Moik, 1980).
Figure 1.0: Point Spread Function
With this criteria the model assumed that the variation of reflectance values of homogeneous areas can be described by a Gaussian function.
When a linear feature or point feature is inside the homogeneous area, the mean of the reflectance valuees of homogeneous area can be described by a gaussia function . when a liner feature or point feature isinside the homogeneous area the mean of the reflection value within a window will be changed. Associated variance will be changed considerably with respect to the change of the mean value.
According to that, heterogeneity can be detected by considering the variation of the coefficient variation within a small window. It should be noted that for small window the statistics is not as valid as it is for large window. But large window do disappear small features.
Procedures for Testing the Model
Two procedures are adopted to test the model.
Procedure (i) :
In this procedure, the variation of the reflectance value in a neighborhood around the center pixel of sub image is used to determine the homogeneity of an area. For that, the center pixel value is replaced by the variance of sub image. The existence of a heterogeneity is indicated by the higher value of the center pixel in the output image associated with the higher variance of the sub image.
The selected window size for sub image be small in order to find the significant variation of the variance due to thin
lines or points.
The algorithm is :
Begin
For every (i,j) position on the original image Do
Take the 3x3 surrounding sub image;
Compute the variance of sub image;
var-image (i,j)=variance of sub image;
End.
By slicing the variance-image at a certain threshold, only the heterogeneous areas (points, lines, edges) are detected.
The algorithm for the thresholding is :
Begin
For every (i,j) position on the var-image Do
If var-image (i,j) >= threshold;
Then binary (i,j) = I; (points, lines, edges)
Else binary (i,j) = 0;
End.
Procedure (ii) :
It considers the rank order of a sub image (3x3). The presence of homogenous areas assume the rank 2 and rank 8 falling within the probability range of 5% and 95% in the Guassian distribution. Therefore, the range between rank 2 and rank 5 is equal to the that of rank 5 and rank 8. In the presence of heterogeneity, two ranges are not equal.
The algorithm is :
Begin
For every (i,j) position on the original image Do
Take the 3x3 surrounding sub image;
Sort its 9 values into the array Rank;
r2_image = Rank(2);
r5_image = Rank(5);
r8_image = Rank(8);
r2_5_image = Rank(5) -Rank(2);
r5_8_image = Rank(8) -Rank(5);
Hetro_image = r2_5_image- r5_8_image;
End.
Conclusion
This study compared extracted features with the information content of the map at 1 : 50000 scale. This comparison shows that main roads and secondary roads can be detected successfully with SPOT multispectral images. However, resolution is important since it was not possible to identify tracks and unpaved roads with low resolution.
The experimental results shows positive indications for
updating main roads and secondary roads in road network at
1 : 50000 scale.
References
-
Moik. J (1980). "Digital Processing of Remotely-Sensed Images", NASA special publication 431, Washinton DC.
- Paul M. Mather (1987). "Computer Processing of Remotely-Sensed Images", John Wiley & sons, ISBN 047190648 4,Chapter 2.
- Parakum Shantha K. D. (1992). "Use of Radar Imagery for Updating Terrain Information in Topographic Databases", Dissertation, ITC, The Netherlands.
