Generation Of Dense Urban Features Using Conditional GAN

Abstract This paper discusses the use of conditional Generative Adversarial Networks (GANs) to generate dense urban features in satellite images and evaluate their effectiveness in semantic segmentation tasks. High-resolution true-color satellite imagery of Mumbai, obtained from Pleiades-1A at a 0.5m resolution, is utilized for the study. The proposed Multiple Discriminator pix2pix (MD-pix2pix) model, which employs multiple discriminators and a modified training procedure, is introduced to generate realistic satellite images. The performance of the MD-pix2pix model is compared to the traditional pix2pix model using a mix dataset of real and generated satellite images. The synthesized satellite images are assessed for their effectiveness in semantic segmentation tasks using various CNN models, including VGG16-UNet, MobileNetV2-UNet, and DeepLabV3+. The study aims to overcome the limitations of existing datasets that do not include informal settlements, such as slum areas, which are common in many cities. The results indicate that the MD-pix2pix model produces more realistic satellite images with greater variability in vegetation type, slum arrangements, and built-ups than the traditional pix2pix model. The synthetically generated satellite images are also effective in semantic segmentation tasks, with better segmentation accuracy achieved using the MD-pix2pix generated images for computationally less expensive architectures such as MobileNetV2-UNet. This study highlights the potential of GANs to generate realistic satellite images for urban feature mapping and monitoring applications.