48.

[Cover] [Contents] [Index]

Page 140

3.5.1. ART_a then starts again to search for another winning neurone, which is again tested using Equation (3.55). Such a process is called match tracking. If the winning neurone in ART_a satisfies Equation (3.55), the weights in ART_a are adjusted as shown by Equation (3.52) except that is replaced by , and the weights between F2_a and F_ab are changed to . The vigilance ρ^a is then set to its initial value before the next training pair is presented. Note that in both ART1 and fuzzy ART systems, the vigilance ρ is a fixed parameter. However, in ARTMAP and fuzzy ARTMAP, the vigilance parameter ρ^a is dynamically changed in order to perform match tracking.

3.6 Neural networks in remote sensing image classification

Over the past decade, the use of neural networks for classifying remotely sensed imagery has developed rapidly, mainly because neural network classifiers are believed to out-perform standard statistical classifiers, such as maximum likelihood. The lack of assumptions concerning data distribution is another factor that make neural networks more attractive than statistical classifiers, especially when the size of training data is limited so that adequate estimates of statistical parameters are difficult to obtain.

3.6.1 An overview

The most popular neural network classifier in remote sensing is the multilayer perceptron. Many classification experiments using the multilayer perceptron are found in the literature. Good surveys are provided by Paola and Schowengerdt (1995), Kanellopoulos et al. (1997), and Atkinson and Tatnall (1997). Kanellopoulos et al. (1992) conducted a 20-class classification experiment on SPOT high resolution visible (HRV) imagery, and the result is satisfactory. Cloud identification is reported by Lee et al. (1990) and Welch et al. (1992); classification using synthetic aperture radar (SAR) is described by Decatur (1989), and a comparison between statistical classifier and multilayer perceptron is provided by Benediktsson et al. (1990). Note that all the above classification experiments used the multilayer perceptron network. The same classification can also be carried out by other main types of neural networks such as SOM, and fuzzy ARTMAP system, though the results might reveal some variation (a comparative study is performed later). In addition to the examples cited above, multilayer perceptrons have also been used for other kinds of classification purposes. For example, Foody et al. (1997) use a multilayer perceptron to explore the subpixel mixture problem, and Jin and Liu (1997) estimate biomass from microwave imagery. A method for forest stand parameter estimation is demonstrated by Wang and Dong (1997).

[Cover] [Contents] [Index]