An Supervised Artificial Neural Network Method for Sattelite Image Segmentation

Image segmentation is an important step in image processing (image processing). The main purpose of segmentation is to simplify and or to change the representation of an image into a form that is easier to analyze. Already there are several methods of image segmentation are found, but most of these methods are not suitable for satellite imagery and -Method method requires a knowledge of the initial (a priori knowledge). To overcome these problems, a satellite image segmentation method is developed using an artificial neural network method without learning (unsupervised) called Kohonen's self-organizing map and threshold technique.

Self-organizing map (SOM) is used to adjust the gray level of the pixel based on the value of the color 3-Dimensional (multiple bands) into a 2-Dimensional. In SOM, the input signal of n-tuple and group-kelompoksebanyak m clusters. Each input is connected to all the units are full. Weights obtained by random and small value that will determine the status of the end (final state). The tissue was prepared by orthogonal grid cluster units (neurons), each of which is associated with three internal weights of three layers of satellite imagery. At each step in the training phase, the cluster unit with very coco weight k with the input pattern will be chosen to be the value of the winner (the Winner) by using the minimum euclidean distance (see formula).

Then a threshold technique is used to group (cluster) of the image into separate areas, this method is called TSOM. The study was conducted on two different satellite images prove that the stability, homogeneity, and efficiency of the method compared with iterative methods TSOM self -organizing the data analysis (ISODATA). The stability and homogeneity of the two methods is determined using the procedure selected based on functional models.

Advantages
1 Does not require a prior knowledge or learning (unsupervised).
2 More stable than the ISODATA.
3 two times faster than the ISODATA when running.

Disadvantages
The efficiency of the method depends on the number of iterations and the threshold value is selected manually.