49.

[Cover] [Contents] [Index]

Page 141

The multilayer perceptron can also be used to detect line features on SPOT panchromatic images. Each training pattern is formed by the nine pixel values falling within a 3×3 window. We selected thirty-one training patterns and used a 9|30|16|2 multilayer perceptron for training. The results are shown in Figure 3.19. Although some line fragments are present on the image, the overall performance seems to be moderately successful.

In comparison with the multilayer perceptron, the Kohonen SOM is not used as frequently for classifying remotely sensed imagery. Besides its use in image classification, SOM can be used for other purposes, such as input data dimension reduction. For example, the number of input features could be very large in a remote sensing classification exercise. The number of input features could be reduced using the divergence index or B distance (Chapter 2), which aim to select the ‘best’ feature subset. An alternative is to use a SOM to perform a mapping for different multidimensional feature groups, and then use the resultant features for classification. Note that, by means of SOM, reduction in the number of input channels and grey levels can be done simultaneously. For instance, if the number of input features is four, each with 256 grey levels, the user may define four neurones in the input layer and 8×8 (= 64) neurones in the output layer. The SOM will output a single image with 64 grey levels.

Although Hopfield networks are mainly exploited for auto-associative memory, a new direction of investigation for temporal feature tracking has emerged in recent years (Bersini et al., 1994; Cóté and Tatnall, 1997). In remote sensing, there is a special need for matching features that vary over time, such as the automatic monitoring of the direction of cloud rotation or iceberg movement, or identifying the same point within multitemporal images in order to automatically construct ground control points (GCPs). The use of a Hopfield network to perform such kinds of operation relies on the definition of a suitable energy function. Studies of multitemporal feature tracking based on remote sensing imagery are given by Cóté and Tatnall (1997) and Lewis et al. (1997).

ART systems, especially ART2 and fuzzy ART, can be useful in performing unsupervised classification on remotely sensed imagery. As described in Section 3.5, the user does not need to decide how many clusters are to be generated; rather, one has to define a vigilance parameter to control the formation of clusters. Higher vigilance levels result in fine clusters, while lower vigilance levels generate coarser clusters. A range of vigilance parameters (from low to high) can be used to perform a hierarchical unsupervised classification.

Figure 3.20 illustrates the effects of the vigilance parameter on clustering performance. We use a fuzzy ART neural network, and adopt the fast learning strategy (i.e. learning rate β=1). The test images are three-band SPOT HRV images, each containing 128×128 pixels. The vigilance parameter was set to values of 0.6, 0.7, 0.8 and 0.9, respectively, and the

[Cover] [Contents] [Index]