Semantic Segmentation of Urban Area using Pix2Pix Generative Adversarial Networks

Land cover analysis using satellite images is essential for understanding how urbanization has affected a particular geographical area. In this paper we present semantic segmentation of Urban Areas using Pix2Pix Generative Adversarial Network (GAN). The GAN architecture proposed in this paper uses modified UNet as a generator and PatchGAN as a discriminator. We have trained Pix2Pix GAN which is trained using a set of two images (input and target) as input to generate the images which look like the target images. The dataset is generated using images taken from Google Earth. Comparison is made among different approaches namely Autoencoder, UNet and our Pix2Pix GAN. The results are evaluated and compared using performance metrics like Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index (SSIM). The experimental results indicate that the proposed model outperforms the other approaches giving a PSNR of 48.26 dB and SSIM of 0.97. The performance evaluation and visual analysis both conclude that proposed approach is robust and accurate for semantic segmentation of urban area.