Image Analysis of Remote Sensing Data Integrating
Spectral and Spatial Features of Objects
3. Road detction from landsata tm images:
The road network development in the forest indicates an expansion of human activities, and it is highly related to the man-made disturbances to forest ecology and forest growth. Therefore, monitoring of these changes from space is one of important applications for remote sensing.
Forest clear-cut patches appear as area-like objects in remote sensing, and it has been investigated in the previous work (J. Iisaka and et al.1995), here the line-like objects are focussed.
The following sections illustrate some direct application of the pixel swapping method to analyse remote sensing data integrating spatial and spectral information of terrain objects.
3.1. Automated line-like object detection.
Image slicing is the most popular method to extract specific objects for many image-processing applications. This threshold value is ordinary determined at locations of peaks or valleys of the histogram of pixel values distribution, or interactively by operators of image analysis. As long as the targets have specific pixel value range and are well separated in the histogram, this method extracts the target objects very well. With multispectral images, the target objects are well identified based on the spectral similarity in spectral space, and the threshold is selected from spectral similarity.
Unfortunately, most terrain covers observed in conventional remote sensing such, as Landsat TM data are not well separated in spectral space except few objects like water. Therefore, simple threshold values are difficult to select solely from the histograms.
Problem are not only separability among objects, but some objects may have several threshold levels, because the radiometric values are affected by the objects of their neighbour or background objects in that sensor resolution cell.
It is assumed that the roads and trails appear as line-like objects and brighter than other object in a scene.
For simplicity, TM brightness image is used as the roads and trails mostly appear as bright objects in a TM scene. The brightness image is created by combining the TM visible bands, TM band 1, TM band 2 and TM band 3, which is illustrated in Figure 2.Figure 2 TM Brightness
Figure 2. Image of Study Area:
Watershed of Greater Victoria, BC, Canda
Figure 3: The Result of Line-like Object Detection
This brightness image is sliced at various thresholds starting from the maximum pixel value in the scene decreasing it with a fixed interval. With the highest threshold, no objects are extracted. With the intermediate threshold between the maximum and minimum pixel values, some objects appear as point-like objects, region like object or line-like objects. Figure 3 shows the objects extracted at the threshold that generates the maximum number of pixels for the line-like patterns.
The first experiment was conducted as follows: 1) Create a series of binary images sliced at every grey level. 2) Each binary image is processed by the pixel swapping method, and each pixel is labelled as a point, a line, or a region. 3) Then, extract only the objects labelled as line-like objects and store these results. 4) Repeat the above procedure to the image of the next slice level. 5) The result image for line-like objects is "ORed" with the previous results.
As seen in the Figure 3, many shorter line-segments are detected in the vegetated areas, and they might be the features generated by the topographic effects or tree height variability in the forest.
3.2 MODVI and Vegetation Mask
These small line-like features in the forest can be masked out using vegetation indices derived from remote sensing data.
Although the NDVI (Normalised Difference Vegetation Index: C. Tucker, 1979) is widely used as a vegetation index, it is not so reliable against radiometric environment of remote sensing. In this experiment, the MODVI (The modified Vegetation Index) is adopted in this experiment, as the MODVI is less affected by the atmospheric and solar illumination conditions. (J. Iisaka and et al. 1999, and J. Iisaka, 2000)
The MODVI is estimated by the following equation.
Here, W (W3, W4) is estimated from the spectral values of darkest objects in TM band 3 and band 4 images, and D3 and D4 is the pixel values of TM band 3 and 4, respectively.
3.3 Fractal Based line-like Object Detection.
It is more convenient if an algorithm is able to identify the line-like object independent to the pixel values or contrast among neighbour pixels. Most of objects observed in nature are not so simple like straight lines or ploy-lines. Rather, they are characterised by fractal dimensions. As the fractal dimension of line-like objects never exceeds 2, no matter how complex they are, fractal measures can be applied to delineate line-like objects.
Let i and j be the indices of scales (the swapping window sizes) , and denote the results of pixel swapping as Si and Sj. The fractal dimension (local fractal dimension) D can be estimated from the following equation.(J. Iisaka, 1998,1999)
In Figure 4, each pixel represents the local fractal dimension (windows sizes were 3x3 and 5x 5).
Therefore, it is necessary to threshold the fractal image between "1 or greater than 1" and "less than 2".
Figure 4 Fractal Image derived from the rightness
image (see Figure 2.)
Figure 5 Road dectected in less vegetated areas
The Figure 5 is the result of roads in less vegetated areas derived using the TM data, and overlaid with TM false colour image of the study area.
