Recently, support vector machines (SVMs) have been widely used with success for hyperspectral image classification. Indeed, a SVM-based classifier makes it possible to achieve more accurate classification under the condition that a small number of training samples are available. However, there are several issues that need to be considered and investigated before SVMs become operational in remote sensing applications. As the kernel functions are the backbone of SVMs, the most important problem is how to determine the kernel function and the corresponding parameters. A conventional method for setting the kernel parameters is the grid search method. Throughout the training process, classifiers are trained with different kernel parameters, and only one of the classifiers is required for the testing process. This makes the training process time-consuming. As a solution, we propose a new method using the intercluster distances in the feature spaces to set the kernel parameters. We propose a multistep algorithm based on the SVMs in a similarity space. These techniques define and use the spectral similarity measures. In this new space, the pixels corresponding to the class-of-interest will obtain similar values. This space can be seen as a new feature space with a relatively low dimension. For the first step, hyperspectral data are projected onto this similarity space. The second step concerns tuning the SVM parameters to separate the class-of-interest pixels from background pixels. The results demonstrate that the proposed algorithm is efficient and reliable for hyperspectral image classification purposes. Furthermore, the experiment’s results showed that the intercluster distance could choose proper kernel parameters with which the testing accuracy of trained SVMs is determined by the standard ones, and the training time could be significantly shortened.