Comparative Study Between Support Vector Machines And Neural Networks For Lithological Discrimination Using Hyper spectral Data

Remote sensing hyper spectral data has many applications especially in the field of , earth science. Utilization of this technology has shown a rapid increase in many areas of economic and scientific significance. Hyper spectral sensors capture the detailed spectral signatures that uniquely characterize a great number of diverse surface materials. Classification, clustering, and visualization of these very high dimensional signatures need untraditional methods. Different approaches for spectral image interpretation have been studied using Artificial Neural Networks (ANNs) and Support Vector Machines (SVM) to meet the challenge of high dimensionality. The study used SVMs for geological mapping of hyper spectral imagery at Abu Zenima area, western Sinai, Egypt, the hyper spectral data has been captured in 2003 by Hyperion instrument on the United States Geological survey (USGS) Earth Observing 1 (EO-I) satellite. Precisely the study compares between the use of SVMs and a neural network built on the concept of SVMs, this network uses the Kernel-Adatron algorithm with the Gaussian kernel for the process of training. The SVMs also uses the Gaussian kernel with different bandwidths to enhance the performance of the interpretation process; the results are compared in details. The Neural Network was trained with four data sets, the first consists of 11310 samples, gives recognition rate of 84%, the second has 22620 samples, recognition rate was 91.5%; the third has 33930 samples, recognition rate was 94.6%; finally the fourth has 45240 samples, recognition rate of 99.2%. The previous results fall in comparison with the results of SVMs which use two algorithms for training the first is the one against one algorithm which gave a recognition rate of 84% for the first data set, a recognition rate of 76.9% for the second data set, a recognition rate of 95.2% for the third one and 98.5% for the fourth one. and the other is one against many algorithms which gave a recognition rate of 84% for the first data set, a recognition rate of n.3% for the first data set, recognition rate of 94.6% for the second one and 98.5% for the third one