Urban Surface Simulation Through Image-to-Image Translation Deep Learning Algorithm using Optical Aerial Imagery

Digital Surface Model (DSM) provides the detailed structure and geometry of an urban environment. This paper proposes an approach of using a type of image-to-image translation deep learning model called cycle consistent adversarial networks for reconstructing DSM from monocular aerial imagery. The cycleGAN architecture consisted of two generators with an encoder-decoder network with skip connections and two discriminators that punishes structures at the scale of patches. The cycleGAN objective function was adapted for training on paired images. The evaluation was performed using mean square error (MSE) and zero normalized cross-correlation (ZNCC) for errors in reconstruction. cGAN model was considered as a baseline model for comparison of the proposed approach. The results using the proposed approach confirmed a higher reconstruction accuracy than previous studies that utilized conditional GAN.