A deep affine registration framework for SAR and optical remote sensing images using SE-enhanced siamese networks

Image registration of SAR and optical images is an important task in many remote sensing applications, such as military surveillance, agriculture, change detection, and environmental monitoring. However, accurate alignment between different modalities remains challenging because of speckle noise, radiometric differences, and geometric distortions. To address these challenges, this paper presents a deep affine registration framework based on an SE-enhanced Siamese VGG16 network and an Affine Regression Network (ARN). The Siamese network extracts features from SAR and optical image pairs, whereas the Squeeze-and-Excitation (SE) blocks strengthen channel responses and reduce the effect of noise and modality differences. The ARN directly estimates translation, rotation, scale, and shear parameters from the extracted features. The proposed framework was evaluated on the SEN1-2 and CNAM-CD datasets, which contain diverse SAR–optical scenes with different imaging characteristics. Experimental results show that the proposed SE-VGG16+ARN framework achieves lower RMSE and higher registration accuracy than the baseline VGG16+ARN model and conventional registration methods. Complexity analysis also shows that the proposed method achieves accurate geometric alignment with competitive inference time. The experimental results demonstrate that our method provides accurate and robust SAR–optical image registration.