Classification of satellite sensed data using genetically optimized auto-associative cellular neural networks

This paper presents the use of Cellular Neural Networks (CNNs) in modeling and classification of land cover/use features based on images acquired through remote sensing methods. Due to their strong local interconnections and simplicity, CNNs are well suited to spatio-temporal applications and have since inception captured the attention of researchers in this field. This paper evaluates the use of associative memories to extract and store features of images acquired via satelite sensing. It is proposed that using Genetic Algorithms (GAs), associative memories more capable of extracting details of a land feature can be achieved in the CNN model.This method relies on a better selection of the training set and the initial filtering process to remove noise on the images. CNN-based diffusion is applied to achieve an average grey-scale level for all images.Using GAs and training samples, a CNN autoassociative template is automatically designed to store features for each class. The templates were applied to the image representing the study area to extract the features belonging to the class they represent. It is concluded that a 3 X 3 CNN template has enough memory to store features of a class. It was found that the pixels on the boundaries between classes are un-classifiable.