A Modified DDN Architecture for Enhanced Change Detection in SAR Images

The field of change detection in synthetic aperture radar (SAR) images has emerged as a critical area of research with diverse applications. This work proposes an innovative approach by integrating convolutional neural networks (CNNs) within a modified dual-domain network (MDDN), simultaneously leveraging both spatial and frequency domain information. Leveraging the multi region convolution module (MRC) in the spatial domain allows the model to adapt and learn from different features present in these diverse regions. To further enhance the efficiency of the proposed approach, the adaptive discrete cosine transform (ADCT) is used in the frequency domain, which plays a pivotal role in selecting critical components. The ADCT incorporates an attention gate mechanism which involves using element-wise multiplication to focus on key elements within the information vector. Simulations convey that the proposed method works well on the ‘Ottawa’ and ‘Sulzberger’ dataset, but the ‘Yellow River-A’ and ‘Yellow River-C’ datasets result in suboptimal performance. Also, the MDDN gives better performance for ‘PCC’ and ‘Kappa’ coefficient.