Satellite Image to Map Conversion Using Conditional Adversarial Networks (cGANs)

Abstract

Translate Article

A Deep learning model that turns satellite images into road maps with the help of a Conditional Generative Adversarial Network (cGAN). A U-Net generator is applied in the architecture to draw details from the road and map, and a multi-scale PatchGAN discriminator reviews the results at several resolutions to enhance both the fineness and accuracy. Adversarial loss works on realism, L1 loss helps with pixel by pixel, and SSIM loss protects the structural elements. The model is designed using the Massachusetts Roads Dataset and it has attained 95% pixel accuracy, along with impressive accuracy and structure similarity. The proposed model provides road masks that are clearer and more reliable, regardless of environmental complexity or lighting conditions. Using this strategy, it is possible to skip manual mapping and support applications in urban planning, navigation, and responding to disasters. In the future, the optimal solution might involve trimming the model's size to make it work more quickly.