The training data selection is an important and operative step in the classification and change detection procedure from remote sensing images, which needs to be provided with high sensitivity. These samples are often determined by the human factor, which is a time-consuming process and prone to high error. Old maps can be a valuable source of information for selecting and preparing training samples. If these samples are accurately refined, they can speed up, facilitate and also increase the accuracy of the change detection process. The main innovation of the present paper is the diligence in the sampling process, which has been made imaginable by proposing a model based on the chi-squared statistical test and k-means clustering. This method, while using Chi-square statistical test, tries to select pure training samples, by selecting samples that are close to the centers of internal clusters in each class with multiple k-means clustering that takes into account the internal spectral diversity of classes. In this method, the spectral and the first and second-order of co-occurrence matrix are extracted and used in the support vector machine (SVM) classification process. Furthermore, the feature selection and SVM parameters have been optimized by the genetic algorithm to more improve the classification and change detection accuracy. For implementation, bitemporal satellite images at 2011 and 2015 and land use map of 2009 related to the Shiraz has been employed. Using the proposed method led to update the thematic maps of the study area with an overall accuracy of 98.72% and 94.57%, and a from-to change map. Experimental results showed that the refinement process of the training samples improves the results of the 2011 image classification (increasing the kappa coefficient from 65% to 87% and increasing the overall accuracy from 73% to 91%) as well as the 2015 image (increasing the overall accuracy from 69% to 86.32% and Kappa coefficient has been increased from 59% to 80.48%).