Keywords: TM images analysis, object recognition, spatial information extraction, application
of fractal dimension, pixel swapping, image analysis of remote sensing
Abstract Digital image analysis techniques have been and are being used widely in remote sensing assuming that each terrain surface category is characterised with spectral signature observed by remote sensors.
Even with the remote sensing images of medium coarse resolution such as Landsat TM data, integration of spatial information is expected to assist and to improve the image analysis of remote sensing data.
This paper will describe and demonstrate a method, named as" Pixel swapping", to analyse remote sensing images in a unified way, integrating spatial information extraction as well as spectral information.
After a brief introduction of the method, some application to identify terrain objects like roads in the forest are demonstrated utilising the spatial information of objects.
1. Introduction
With remote sensing images such as LANDSAT TM data, most of digital analyses are applied through pixel-by pixel basis, assuming that each ground cover objects has their specific spectral characteristics.
The conventional digital image processing techniques, based on the field theoretic approach, have many advantages in numeric and analytical calculation such as filtering the signals and classifying terrain covers using multi-spectral characteristics of objects, but they are not so capable to handle spatial and geometrical information.(J. Richard, 1994, J. Jensen, 1995).
On the other hand, the mathematical morphology, based on a set theoretical approach, provides better capability with morphological processing of objects, but week in numeric and quantitative treatment of objects. Thus, more powerful functions to handle spatial feature extraction for remote sensing are required.
2. Spatial Information Extraction For Remote Sensing
As an extension of conventional image processing and mathematical morphology, It has been proposed a method to process spatial information in an image.(J.Iisaka,1989, 1998, 1999, 2000, and J. Iisaka, et al, 1993 and 1994)
2.1 Image Computing through Pixel Swapping.
An image is defined as a function image intensity field I with parameters of a coordinate pair, x and y as:
I= f (x, y) (2.1)
For a digital image, the set of co-ordinate value, X, is a finite set, and the function (2.1) is a projection of a set X to an element of a pixel value set, a, and can be denoted as:
I= {(x,a(x):x
eX)} (2.2)
Here, a is a set of the pixel values
This equation can be rewritten again using two points, dividing it into two terms as
a(xi)=[{a(xi)+a(xj)}/2
+{a(xi)-a(xj)}/2] (2.3)
This means that an image is composed with two components, one represents an additive term between a pixel at the co-ordinate and another pixel at the co-ordinate and the other represents a subtractive term between them.
Let's demote the first term as
I+={(x,{a(xi)+a(xj)}/2:x
eX)}
and the second term as
I-={(x,{a(xi)-a(xj)}/2:x
eX)} (2.4)
Then, an image, I, is presented as
I=I
++I-
(2.5)
The selection of a pixel pair between pixel i and j is arbitrary. A homogenous flat image can be identified if the image is not affected against any pair of pixels in that image, and if the image is not affected against any pixel pair arranged in the horizontal direction, but affected against other directions, the image may include horizontal objects. In this way, a strategy for pixel pair selection defines the spatial feature of interest.
Foe simplicity, let's select the counter pixel j to be paired with a focusing pixel i from the adjacent eight neighbour pixels and take a sum of them.
Here, n is the number of pairs.
The first term corresponds to a weighted image smoothing function, and the second terms is same as the eight neighbour Laplacian edge detection kernel, which are ordinary treated separately through conventional image processing. As the operations described in the above have dilation (or erosion) effects to the objects, it would be necessary to eliminate these effects.
The simplest and most fundamental application it to identify the objects in terms of spatial entities as point-like, line-like and region-like objects. Other immediate application of the method is to detect the spatial feature points such as line-start and end points, vertices or corner points, or branch-points and line crossing points.
Figure 1. A test image with ideal shapes and the result of pixel swapping , using a 7x7 circular kernel. The image is divided into inner regions (yellow), boundaries (red), and lines (blue).
