Convolutional neural network architecture for digital surface model estimation from single remote sensing image

A convolutional neural network (CNN) architecture has been proposed for estimating the digital surface model (DSM) from a single airborne or spaceborne image, which is inherently an ambiguous and illposed problem. Deriving the three-dimensional information and reconstructing the geometry of a surface from a monocular image require a deep network that has the ability to extract the local and global characteristics of the surface. In order to address this challenging issue, a deep CNN with residual blocks is employed as a downsampling part of the network, and an upsampling procedure is presented for improving the output accuracy. Moreover, an approach is proposed for connecting the estimated DSM patches and generating a seamless continuous surface. In order to assess the proposed methodologies, scenarios are designed and implemented in various datasets. The final results show that the root mean square error (RMSE) of the proposed approach, when the training and testing images are selected from remote sensing images, is about 3.5 m. In addition, evaluating the capability of the proposed approach for depth estimation using the terrestrial images of outdoor scenes reports about 4 m for RMSE, which is better than the other methods in the literature and demonstrates the promising performance of the proposed approach.