Optical and SAR images-based image translation for change detection using generative adversarial network (GAN).

Monitoring a specific area to analyze a continuous change has become more accessible by using optical images in remote sensing technology. However, several natural and artificial aspects such as fog and air pollution make it difficult to extract correct geometric information. To overcome the limitation of optical images, Synthetic Aperture Radar (SAR) images can be used to access more accurate information with respect to the targeted area. In this manner, optical and SAR images can be utilized together to detect the scale of change even in bad weather conditions. To process optical and SAR images, an image translation process-oriented Deep Adaptation-based Change Detection Technique (DACDT) is proposed. An optimized U-Net++ model is proposed that helps to improve the global and regional impacts of the images. Moreover, a multi-scale loss function is utilized to access the features of different dimensions. In this manner, the final change maps are generated by transferring the features of optical images to the SAR images for better change analysis. The prediction performance of the proposed approach is evaluated on four different datasets such as Gloucester I, Shuguang Village, Gloucester-II, and California. The calculated outcomes define the prediction performance of the proposed solution by registering the accuracy of 98.67%, 99.77%, 97.68%, and 98.87%, respectively. [ABSTRACT FROM AUTHOR]