Once the imagery has been orthorectified and placed in its correct geographical location, image preparation is the next step. Image preparation involves applying a frost filter to the imagery, border and frame extraction and calculating the NDVI. These steps create additional layers necessary to assist the segmentation process and ultimately the classification.
3.3.1 Frost filter application
Following orthorectification of the imagery, the next step is to apply a frost filter to the imagery. A frost filter is primarily used on radar imagery to remove high-frequency noise while preserving high-frequency features (edges) (PCI Geomatica 2010). Since SPOT 5 imagery contains a large number of pixels due to its high resolution, it would take a significant amount of time for eCognition to acknowledge and process each pixel during the segmentation process. Consequently, a filter must be applied to smooth pixels but leave edges and sharp features unaltered. Although a frost filter is primarily applied to radar imagery, in imagery such as SPOT 5 it can be used instead of a Lee filter for suppressing speckle, while still preserving edges (Liu & Jezek 2004). A speckle filtering consists of moving a kernel over each pixel in the image and applying a mathematical calculation using the pixel values under the kernel and replacing the central pixel with the calculated value. The kernel is moved along the image one pixel at a time until the entire image has been covered. By applying the filter, a smoothing effect is achieved and the visual appearance of the speckle is reduced (Mansourpour, Rajabi & Blais 2006). Because the edges have been preserved, the next preparation step is to apply border and frame extraction to the image that has been smoothed. Segmentation in eCognition groups pixels according to weights assigned to the process in keeping with spectral, textural and shape characteristics. To ensure that the edges of features in an image are enhanced, an edge detection of these images must be done. Once the frost filter has been applied, the extraction of borders is done through processes conducted in Erdas Imagine. Erdas
Imagine‟s model maker is a useful tool for conducting individualised processes on the imagery. The appropriate exact models designed in this study are discussed in the sections devoted to border and frame extraction below. There are two main forms of border extraction, named border and frame here. The filters applied to extract border and frame are separate additional processes described graphically by De Kok & Wezyk (2005).
3.3.2 Border extraction
For borders to be extracted successfully, a maximum amount of contrast between areas of differing spectral characteristics is a prerequisite. These contrasts facilitate the creation of borders through selective filters applied in models in the Erdas Imagine platform. There are many filters that can be used to remove noise, particularly from imagery with a high resolution. The principal way of determining which filter is best for an image is to apply each one experimentally. A visual inspection revealed that the most suitable filter is a Lee-sigma filter found in the radar interpretation area of Erdas Imagine. A Lee-sigma filter uses the statistical distribution of the digital number (DN) values in the moving window to estimate what the pixel of interest should be. Speckle in imaging radar can be mathematically modelled as a multiplicative noise with a mean of 1 (Erdas Imagine Geospatial Imaging 2009). Applied with a coefficient variance of not more than 5, noise was removed rendering the model to be applicable to the image. The creation of the model is shown graphically in Figure 3.5.
Figure 3.5 Model to obtain border of image
X ls
X – ls x 10
With a filter applied to the image, the border can be extracted from it by means of model maker in Erdas Imagine. As shown in Figure 3.5, there are two input parameters {X} and {ls}. The first input is the original image and the second the image the Lee-sigma filter was applied to. The function in the model is the Lee-sigma filter (ls) image subtracted from the original image and then multiplied by 10 (De Kok & Wezyk 2005). The output is an image from which edges have been extracted and the surrounding areas darkened so that there is a stark contrast between the two, enabling the edges to be seen very clearly. The result, illustrated in Figure 3.6, shows borders that have been extracted successfully. The borders of the orchards are clearly defined as well as the lines of the road.
Figure 3.6 Result of border extraction for classification
The extraction of frame will be the continuing step from generating the border layer and is discussed below.
3.3.3 Frame extraction
The second phase of extracting boundary lines is to calculate frame. Frame is the inverse of border and thus the values of the pixels are offset outside the boundary lines of border. That is, the values are the opposite of border where, for example if the value of a pixel lying on the edge of a feature in the border image is 1 (an edge, i.e. white) and another pixel is 5 (no edge, i.e. black), the image containing frame will have the inverse, thus value 1 = 5 will show the value 5 = 1.
The first step to achieving frame is to invert the histogram of the image in Erdas Imagine‟s model maker. As seen in Figure 3.7A, the input is the original image X. The function is to multiply X by (- 1) (De Kok & Wezyk 2005). The procedure multiplies all image values by -1 and inverts the image‟s histogram to a negative. Inverting the histogram of the image creates the first step needed to calculate frame. Following the inversion process, the Lee-sigma filter can be applied to the image again, but this time it is done on the negative image created and not on the original image as done in border extraction. Again, the coefficient of variance must exceed five.
A B
Figure 3.7 Inverting histogram model (A) and model to obtain frame (B)
When the filter has been applied to the negative image, it can then be applied in the same model used for border extraction. In Figure 3.7B the original image is the negatively derived image and the .ls image is the negative image that has had the Lee-sigma filter applied to it. The resulting image is the inverse of border. Figure 3.8 displays the edges that are now the inverse of border. A comparative analysis of Figures 3.6 and 3.8 show that the edges are the opposite. Where the edges in the border image are highlighted white and lie within the feature boundary, the edges of the frame image are highlighted in white outside the feature boundary. During the segmentation process in
-X -ls (-X)-(-ls) x10 X X (-1) -X
eCognition, the parameters can be set to segment only the border and frame layers to ensure optimum extraction of features.
Figure 3.8 Result of frame extraction for classification
The border and frame layers are two very important elements used to determine the fine edge boundary lines of certain features. Proceeding from the border and frame extraction is the generation of an NDVI layer from the multispectral band of the SPOT 5 image.
3.3.4 Calculating NDVI
The normalised differentiation vegetation index (NDVI) is a measure of plant vigour in an image. Typically, an area that is irrigated or has high water availability will show plants that are growing and therefore photosynthesising. Using NDVI as an additional layer in the classification will assist in segmenting and later classifying plant groups from either agricultural fields or natural areas. The NDVI is calculated as follows:
RED Red band
This equation is sourced from Carlson & Ripley (1996). NDVI correlates with photosynthetic activity of vegetation and provides an indication of the “greenness” of the vegetation (Defries, Hansen & Townsend 1995). Image values where the vegetation is dense have positive values, whereas areas with less dense vegetation or bare sand or soil show negative values. Assisting with linear features, which include roads, railways and canals as well as other structures such as centre- pivot irrigated agriculture, digitised vector layers are used. The additional raster layers are included in the discussion below.