In the past few decades as a result of urban population, spatial development of urban areas has been growing fast. This has led to some changes in the environment in these areas. Hence, detecting changes in different time periods in urban areas has a great importance. Conventional CD methods partition the observation space linearly or rely on a linear combination of the multitemporal data. As a result, they can be inefficient for images corrupted by either noise or radiometric differences that cannot be normalized. On the other hand, one of the main challenges in the production of maps of changes in urban areas, Constraints on the spectral separation of bare land and built-up area from each other in these areas. Therefore, in this paper, an automatic kernel based change detection method with the ability to use a combination of spectral data and spectral indices have been proposed. First, the spectral index for the separation of classes covering the urban area of multi-temporal images are extracted. In next step, differential image was generated via two approaches in high dimensional Hilbert space. By using change vector analysis and determining automatically a threshold, the pseudo training samples of the change and no-change classes were extracted. These training samples were used for determining the initial value of kernel C-means clustering parameters. Then, an optimizing a cost function with the nature of geometrical and spectral similarity in the kernel space is employed in order to estimate the kernel based C-means clustering’s parameters and to select the precise training samples. These training samples were used to train the kernel based minimum distance (KBMD) classifier. Lastly, the class’s label of each unknown pixel was determined using the KBMD classifier. To assess the accuracy and efficiency of the proposed change detection algorithm, this algorithm were applied on multi-spectral and multi-temporal Landsat 5 TM images of the city of Karaj in 1987 and 2011. Respect to the features used, the sensitivity analysis for proposed method carried out using five different feature sets. In order to assess the performance of the proposed automatic kernel-based CD algorithm in the case of using DFSS (Accuracy: 86.40 and Kappa: 0.83) and DFHS (Accuracy: 85.54 and Kappa: 0.82) differencing methods, we compared this technique with well-known CD methods, namely, the MNF based (Minimum Noise Fraction) CD method (Accuracy: 77.42 and Kappa: 0.76), SAM (Spectral Angle Mapper) CD method (Accuracy: 64.60 and Kappa: 0.60), and simple Image differencing CD method (Accuracy: 73.44 and Kappa: 0.70). The comparative analysis of proposed method and the classical CD techniques show that the accuracy of obtained change map can be considerably improved.