Deep learning-based generation of digital surface model (DSM) using a single high-resolution SAR image

The main goal of this proposal is to use a convolutional neural network to map a single SAR image into an elevation model. A set of RADARSAT-2 images and LIDAR based DSMs are available to train the network. Thus, the trained model can be used later to build DSM of any area on Earth. Currently, several deep learning solutions have been proposed in the literature for remote sensing image processing. These data-driven methods usually define different network architectures while use a similar cost function (usually RMSE). Such methods may reach their limits and show poor performance in complex regions, such as edge and boundary areas. Thus, the idea is to define a multi-objective cost function to achieve the best performance in different regions. The initial solution can be achieved with the network proposed in the following paper.