52.

[Cover] [Contents] [Index]

Page 144

more stable. The same observation is also made by Mannan et al. (1998) who found that the fuzzy ARTMAP out-performs both the multilayer perceptron and maximum likelihood methods, in terms of classifier accuracy. However, one should note that, in order to obtain good results, one might have to test a range of ART model-associated parameters. Other studies using ARTMAP are Borak and Strahler (1999) and Muchoney et al. (2000).

3.6.2 A comparative study

In this section, the classification performance of four types of neural networks (multilayer perceptron, Kohonen SOM, counter-propagation and fuzzy ARTMAP) is compared. The study area is located near Feltwell, Nor-folk, in eastern England. The input data consist of seven ERS-1 three-look radar images obtained for the 1993 crop-growing season (20 April, 9 May, 25 May, 16 June, 29 June, 18 July and 3 August). Plates 1a,b,c show a colour-composite image made from the first three multitemporal images. It is clear that the images are seriously affected by speckle. A Lee filter (Chapter 1) with a window size of 5×5 was applied to each image in order to reduce this speckle effect. The resulting filtered colour-composite image is shown in Plate 1b. The classification experiment is based on seven speckle-filtered images, and seven information classes (crop categories) were chosen for identification. The legend and numbers of test pixels are shown in Table 3.1. The ground reference image is displayed in Plate 1c. A total of 2,297 training samples was selected, using a stratified sampling strategy. The reader is referred to Table 3.1 for details of the number of training pixels for each information class.

A multilayer perceptron with the structure 7|28|48|7 was arbitrarily chosen and trained with learning rate set to 0.2. The network converged after around 5,000 iterations, which took nearly 30 CPU minutes. Three Kohonen SOM networks, with (a) 7×7, (b) 12×12, and (c) 25×25

Table 3.1 Classification legends, the corresponding number of ground truth pixels, and the number of pixels used in training the neural networks

No.	Information classes	No. of ground truth pixels	No. of training pixels
1	Grass	11868	395
2	Winter wheat	34038	309
3	Spring wheat	9774	325
4	Potato	15378	341
5	Sugar beet	30142	287
6	Carrot	10341	344
7	Spring barley	1776	296
Total	—	113317	2297

[Cover] [Contents] [Index]