The original data produced by the Shuttle Radar Topography Mission (SRTM) tend to have an abundance of voids in mountainous areas where the elevation measurements are missing. In this paper, deep learning models are investigated for restoring SRTM data. To this end, we explore generative adversarial nets, which represent one state of the art family of deep learning models. A conditional generative adversarial network (CGAN) is introduced as the baseline method for filling voids in incomplete SRTM data. The problem regarding shadow violation that possibly arises from the CGAN restored data is investigated. To address this deficiency, shadow geometric constraints based on shadow maps of satellite images are devised. In addition, a shadow constrained conditional generative adversarial network (SCGAN), which incorporates the shadow geometric constraints into the CGAN, is developed. Training the SCGAN model requires both the remote sensing observations (i.e., the original incomplete SRTM data and satellite images) and the ground truth data (i.e., the complete SRTM data, which are manually refined from the incomplete SRTM data with the reference of in-situ measurements). The integration of the multi-source training data enables the SCGAN model to be characterized by comprehensive information including both mountain shape variation and mountain shadow geometry. Experimental results validate the superiority of the SCGAN over the comparison methods, i.e., the interpolation, the convolutional neural network (CNN) and the baseline CGAN, in SRTM data restoration.
We explore the use of convolutional neural networks (CNNs) for filling voids in digital elevation models (DEM). We propose a baseline approach using a fully convolutional network to predict complete from incomplete DEMs which is trained in a supervised fashion. We then extend this to a shadow constrained CNN (SCCNN) by introducing additional loss functions that encourage the restored DEM to adhere to geometric constraints implied by cast shadows. At training time, we use automatically extracted cast shadow maps and known sun directions to compute the shadow-based supervisory signal in addition to the direct DEM supervision. At test time, our network directly predicts restored DEMs from an incomplete DEM. One key advantage of our SCCNN model is that it is characterized by both CNN data inference and geometric shadow cues. It thus avoids the data restoration which may violate shadowing conditions. Both our baseline CNN and SCCNN outperform the inverse distance weighting (IWD) based interpolation method, with the shadow supervision enabling SCCNN to obtain the best performance.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.