A Deep Learning Kernel K-Means Method for Change Detection in Satellite Image

Kernel methods are widely used for feature extraction or classification problems because of its advantage due to their good optimization and nonlinear expressive power. Meanwhile, Deep learning technology and related algorithms are the latest trend in vision, speech, audio, and image processing. In this project, a deep-learning based Kernal K-mean method is used to detect changes in bi-temporal satellite images. Nonlinear clustering with the help of deep learning is utilized to partition a pseudo-training set of pixels representing both changed and unchanged areas. Once the optimal clustering is obtained, the learned representatives are used to partition all the pixels of the multitemporal image into the two classes. To optimize the parameters of the kernels, an unsupervised cost function is used. By exploiting the expressiveness of nonlinear kernels with the learning ability of deep networks this project was able to attain an accuracy of around 96 % in average.