Combining The Spectral and Spatial Signature of Information Classes using Artificial Neural Network Based Classifier for Remote Sensing of Spatially Heterogeneous Land-Use/Land Cover System in the Tropics
Methodology
The Artificial Neural Networks
An ANN is a parallel processing method which simulates the brains bio-neurological structure (Shoshany and Guedalia, 1994). Its simplest element, the nodes can be made to store experiential knowledge through the process of task examples. The nodes simulate the functions of the fundamental cells of the brain, the neurons, together with the function of the cells' dendrites and axons, the cells' output and input fibers, respectively. Knowledge of pattern recognition can be modeled within the ANN through training. It is through this method where information patterns will be presented in the input layer of a neural network while the output layer will be "forced" to output the desired results. It is through constant repetition of this process that the network will learn to output the correct solution in response to the presented input signal in the input layer. Once training is completed, knowledge is said to be successfully modeled within the ANN and the networks are now available for use.
In computer terms, the process is called ANN training. The following is an example of a three-layered ANN with one node in each layer.
Clearly, the network is trained to output 1 when the input is 0 and output 0 when the input is 1. In this example, the training patterns are (0,1) and (1,0) representing the inputs and outputs, respectively. Using a particular computing Method, the network can be stimulated to generate outputs (1 and 0) using the training inputs (0 and 1), the weights (W1 and W2) and the nodes' thresholds (N1 and N2). Using this scheme, the network will be repeatedly fed with the training inputs and "forced" to generate the training outputs by adjusting the values of the weights and thresholds. After a certain number of iterations the network started to learn how to output nearer to 1 when the input is 0 and nearer to 0 when the input is 1. The details of the mathematics behind the computations during ANN training are given by Bandibas (1995), Alexander and Morton (1990) and Rumelhart et al. (1996)
Designing ANN to Model the spatial and Spectral Signatures of Information Classes
Based from the aforementioned approach, a bigger ANN can be designed and trained to extract spatial and spectral signatures of information classes in satellite images. Figure 1 shows the structure of the ANN for the spatial and spectral signature extraction of information classes in digital images. The number of ANNs is equal to the number of information classes consider. To extract spatial pattern in the images, a pixel window was defined. The pixel values and their unique spatial arrangement in that windows will be the spatial identity of the central pixel in that window. On the other hand, the spectral information was also incorporated into the spatial signature by using more than one spectral bands during the pixel window definition.
Supervised Training
In this study, training areas were specified during the field survey of the study site. The pixel window was moved over the specified training area for each information class and sample patterns were extracted. A Borland Turbo C++ programming language was specifically written for the sampling of pixels. The program was run on a 486 IBM compatible personal computer.
ANN Training and Classification Strategy
Training the network involves the presentation of the representative patterns of all the information classes to the network. As shown in Figure 1, each class has its corresponding network. Obviously, each of these networks has only one output neuron which was designed to output a continuous value ranging from 0 to 1. During the training, all patterns from al the information classes were presented to each of these networks. If a pattern entered belongs to that network's corresponding class, it is "taught" to output 1, otherwise, it is trained to output 0. After several iterations, the network started to learn to output higher values (closer to 1) if the pattern presented belongs to the current class trained and lower output (closer to o) If the presented pattern belongs to another class. In this study training ended when the networks can classify the training data with an accuracy of more than 95%. This required periodic testing of the networks' performance after a designated number of iterations interval. Testing involved the entry of a training pattern to all the classes' network. The one which gave the highest value (nearest to 1) is the "winner". This means that it was this network's corresponding class where this pattern belongs. This process was repeated using all the training patterns and the classification accuracy was computed. Similar Procedure was followed during the classification. In this case, the pixel window was moved over the multispectral images and the extracted pixel pattern was presented to all the information classes' trained networks. The corresponding class network that gave the highest value is the class label to be assigned to the center pixel of the window.
Figure 1.The Error Black-Propagation ANN Structure for the extraction of the spectral and spatial signature of information classes using 3 satellite bands and a 3 by 3 pixel window with n number of information classes.
