An Unsupervised Neural Network For Change Detection in SAR images

Most existing change detection methods of SAR images are usually sensitive to the speckle noise, and require a large number of labeled pixels obtained through manual annotation in the training phase. In this paper, a new approach has been proposed to improve the accuracy of change detection in Synthetic Aperture Radar (SAR) imagery. The proposed algorithm uses a cluster algorithm to divide pixels into three categories: changed, unchanged and to be determined. Then a small number (200 in our experiments) of changed and unchanged pixels are randomly selected and sent into self-organizing maps (SOM) to generate synthetic training data. The advantage of SOM is that the neurons in SOM are adjusted specifically for pattern categories through competitive, unsupervised or self-organizing learning. Finally, the synthetic training data are used to train a convolutional-wavelet neural network (CWNN). In CWNN, dual-tree complex wavelet transform can significantly reduce the effect of speckle noise. This method is evaluated on two real SAR image data sets. As compared methods, we have considered PCAK means, CWNN, etc. The experimental results confirmed the effectiveness of the proposed approach.