Abstarct: Automatic change detection in Synthetic Aperture Radar (SAR) images is a process in which two registered images from same geographical area and in different times, are being analyzed by computer algorithms developed for this purpose in order to specify changes occurred on surface. Change detection has a key role in updating geo-data which is an essential work in modern surveillance. In this thesis, for the purpose of change detection in SAR images, we have used of a specific sort of deep learning techniques called PCANet combined with a machine vision technique called Saliency Detection. PCANet is capable of learning non-linear relations between correspondent parts in two images, therefore it is, to a high extent, robust against existing noise in these images. Saliency detection has been inspired by biological visual attention system in which there is one or several center of attention. Selection of training samples is of significant importance in order to achieve good precision and efficiency in SAR change detection task. However, these training samples have traditionally been extracting from whole image, leading to much longer training time and lack of classification precision due to unbalanced number of two class’s samples. In order to overcome this challenge, a combination of saliency detection and PCANet is deployed. For enhancement of training samples’ reliability and reducing number of training samples, context-aware saliency detection is used in order to obtain salient regions in image. In this procedure, training samples for PCANet are being extracted only from obtained salient regions which in turn leads to a far fewer number of training and also test samples than previous task. To increase the reliability of training samples even more, the data number of both changed and unchanged classes are set to be equal. Then, the PCANet is trained using extracted training samples and the remained pixels in the salient regions are being classified by trained network. In PCANet learning process, firstly, filters parameters of each laxyer are analytically obtained using training dataset. These filters capture the most discriminative features from the dataset. These features are used in order to train a linear Support Vector Machine which consequently provides the separating hyperplane. Test data is then classified using obtained model and the final change map is achieved as a binary image. Experimental results on four multitemporal SAR datasets has illustrated the superiority of PCAnet combined with Saliency detection over several well-known previously proposed methods in terms of both detection precision and computation time.