Change Detection of SAR images based on Spatial and Frequency Domain Network

Change Detection (CD) is a vital study of the geosphere changes through synthetic aperture radar (SAR) imagery. In order to tackle the CD problems in remote sensing, deep learning is being extensively utilised due to its deep feature representation along with nonlinear problem modeling. In general, most of the CD methods use patch-wise extraction which results in noise in marginal regions of patches and the existing models do not focus on frequency domain. Therefore, to suppress the noise both frequency and spatial domain extraction of features are performed. Alongside to emphasize the central information we use a modified dual-domain network. The feature extraction in the frequency domain is done with the help of Digital Fourier Transform (DFT). In the spatial domain, multi-region convolutional (MRC) modules are used which comprises the following layers: Convolutional layer, batch normalization, Max pooling layer. Moreover, filtering operation is performed before the difference image generation using mean and median filter in order to suppress the speckle noise in SAR images. Finally, training and testing result will deliver the binary change map. The results obtained through the experiment using the proposed method on the satellite imagery depict the effectiveness of the model.