REFINING CLASSIFICATION ACCURACY FOR SATELLITE IMAGES

Abstract

Classification of remotely sensed data, and especially satellite data, has serious

problems that lessen the classification accuracy resulting from conventional classifiers such as neural network or Bay's classifier. The most appearing problem in satellite data is the mixed pixels problem. The subdivision of a scene into discrete pixels acts to average brightnesses over the entire pixel area so different features or landuses may exist in the same pixel area, the pixel reflectance is a weighted average of reflectance's of different features. Another problem is the dependency of the classification on different sources of data rather than the image data. The employed data may be vague, uncertain, or incomplete. Also, it is difficult to prepare a homogeneous training set that is required by conventional classifiers. The structure identification (the identification of the number of nodes and the number of layers in neural networks, the number of rules and the number of variables involved in each rule in fuzzy systems) is a serious problem in classifiers such as neural network which has no direct way to determine the structure information. Other classifiers such as Bay's require some prior knowledge to perform well. To confront the above mentioned problems we make use offuzzy classifiers that can represent vague data, model mixed, easy to extract their structure, and can .employ expert knowledge. The employed fuzzy system can be trained using neural concepts and hence can be called neuro-fuzzy system. Traditional neuro-fuzzy systems such as ANNBFIS system that is trained using gradient method may got stuck in a local minima, may not distribute modeling of the information classes equally over all rules in the rule set, and may take long time to converge or not converge at all. We make use of ANNBFIS model which is optimized using Marquardt method which converges faster than the gradient method, combined with deterministic annealing to ensure global minima, and no redundancy in rules, and the CAR method to ensure faster and better modeling of information classes. The proposed model automatically extracts the structure of the rules using rule splitting method and employs an adaptive learning rate to ensure training stability. The proposed model is tested on a 3 bands (namely band 2, 4, 7) satellite data (Thematic Mapper data) with 9 landuses classes. Results prove an advantage over the traditional neural network with back-propagation, yielding about 6% increase in accuracy, whereas over the basic ANNBFIS, yields about 3% increase in accuracy.