Kernel methods are gaining popularity in image processing applications. The accuracy of feature extraction and classification on image data for a given application is greatly influenced by the choice of kernel function and its associated parameters. As on today there existing no formal methods for selecting the kernel parameters. The objective of the paper is to apply machine learning techniques to arrive at suitable kernel parameters and improvise the accuracy of kernel based object classification problem. The graph cut method with Radial Basis function (RBF) is employed for image segmentation, by energy minimization technique. The region parameters are extracted and applied to machine learning algorithm along with RBF’s parameters. The region is classified to be man made or natural by the algorithm. Upon each iteration using supervised learning method the kernel parameters are adjusted to improve accuracy of classification. Simulation results based on Matlab are verified for Manmade classification for different sets of Synthetic Aperture RADAR (SAR) Images.