From Objects to Pixels
Pixels and their neighbourhood
Contrary to pixel-oriented classification which analyses the pixels in isolation, object-oriented classification looks at the neighbourhood of a pixel. The basic assumption is that the pixel's probability to belong to the same class as its neighbour is very high. Additionally, it is assumed that the pixels of a class stand out from the other classes in terms of their spectral characteristics. The results are coherent areas, the so-called segments, which are separated from other segments by a distinct border. Normally, there is no transitional zone between segments.
Like in pixel-oriented classification, object-oriented classification incorporates unsupervised and supervised classification procedures. The following sections will address two segmentation processes, i.e. edge-based and region-based segmentation, ranking among the unsupervised classification procedures. The supervised classification procedure is introduced by means of hierarchical clustering.
Unsupervised object-oriented classification
The classification process in object-oriented analysis starts with the division of the image at hand to produce coherent areas. Then, those areas are assigned to a class.
The term segmentation describes the division of a satellite image in several homogenous areas. This procedure is similar to human comprehension, because it searches for similar continuous areas and recurring patterns in a satellite image. The automatic procedure producing those segments works according to one of these two principles:
Edge-based segmentation searches for edges in a satellite image. They assume that the areas between the edges are homogenous, meaning, they are the same. The higher the threshold is that determines whether a change in the pixels' values is seen as an edge, the less edges are there in an image and the less different areas will become apparent in the segmented image.
|Step 1: Search and highlight edges.||
Step 2: Highlight homogenous areas, one colour per segment.
Region-based segmentation does not search for differences of neighbouring pixels, but for very similar pixels. If one pixel is very similar to one or more of its neighbours, it will form a segment with these pixels. Borders between segments exist due to too small similarities, defined by a pre-set threshold.
The pixels are examined in relation to their neighbourhood and are aggregated to form homogenous areas.
In the proper meaning of the word, segmentation is a very special kind of classification. Because different segments in different places can belong to the same class as well, we look at the spectral characteristics of the segments after segmentation. The various spectral characteristics of the pixels within a segment can be averaged to compare different segments. If the spectral signatures of two or more segments are similar, they can be assigned to the same class. Either the researcher determines the number of classes or the similarity index value, a threshold which states when the class assignment doesn't seem reasonable any more.
The same threshold (minimum 3 pixels in a segment) results in different segmentations for region-based (right) and edge-based (left) segmentation. The images above show the segmentation result after they have been edited - the segments have been summarized in classes and coloured with the same colours.
Supervised object-oriented classification
As well as unsupervised classification, supervised object-oriented classification segments the image and classifies the segments afterwards.
Here, we have again two possible ways of classifying the segments: The first is a visual determination of the edges between segments and their manual mapping. The second procedure is the determination of criteria which have to be met in order to assign a pixel to a segment. For example, at least two pixels in the neighbourhood have to have the same grey-scale values. Or: The segments can only be separated by streets. In order to use this last statement as a criterion, the street data set is needed.
Having segmented the image, the segments can be used as training areas, just like in the supervised procedure of pixel-oriented classification. In addition to the grey-scale values of the pixels, the texture of the segments can be used as well. The texture is formed by the distribution of the grey-scale values within a segment, e.g. for fields the texture is determined by drills and plant rows. Whenever this texture is observed in combination with the mean grey-scale value, this segment is assigned to the class "fields".
For value X in the green band: Acre
Classification procedures are called "hierarchical" if the researcher defines if-then-rules for image classification. The computer has to learn how to distinguish between water and non-water, for instance. Selecting pixels exemplary for the class water or non-water, the researcher has to train the computer on spectral characteristics. Thereafter, the whole image can be divided into, those two classes and, subsequently, new rules can be established to differentiate within those classes.
The class "field" can be differentiated in fallow land and cultivated area, the class "non-field" in water and land and so forth. Leading to a tree structure, this procedure is also called decision tree classification.
Object-oriented classification segments image data. In contrast to pixel-oriented classification, homogenous areas with sharp borders are produced. Object-oriented classification can either be performed manually or automatically.