A SAR Change Detection Method based on Convolutional Neural Networks

摘要

To reduce the influence of coherent plaque on change detection in synthetic aperture radar (SAR) images and reduce the labor cost of labeled samples, this paper developed an unsupervised SAR change detection method based on hierarchical fuzzy C-means (FCM) and convolutional neural networks. Firstly, the difference between the multitemporal images was calculated by using the neighborhood-based ratio (NR) operator, and the difference image was unsupervised initially segmented into the changed, unchanged and pixels to be determined by using the hierarchical FCM. Then, to solve the problem of sample imbalance during change detection, a data sampling method with frequency-invariant downsampling was proposed. High-confidence changed and unchanged samples were selected for network training. Finally, the trained neural network was used to classify the pixels of the determined to obtain the final change result. Experiments were performed by using SAR images and the results showed that the method was convenient and effective and showed a high accuracy.