Void filling of digital elevation models with a terrain texture learning model based on generative adversarial networks

Abstract

Digital elevation models (DEMs) are an important information source for spatial modeling. However, data voids, which commonly exist in regions with rugged topography, result in incomplete DEMproducts, and thus significantly degradeDEMdata quality. Interpolation methods are commonly used to fill voids of small sizes. For large-scale voids, multi-source fusion is an effective solution. Nevertheless, high-quality auxiliary source information is always diffcult to retrieve in rugged mountainous areas. Thus, the void filling task is still a challenge. In this paper, we proposed a method based on a deep convolutional generative adversarial network (DCGAN) to address the problem of DEM void filling. A terrain texture generation model (TTGM) was constructed based on the DCGAN framework. Elevation, terrain slope, and relief degree composed the samples in the training set to better depict the terrain textural features of the DEM data. Moreover, the resize-convolution was utilized to replace the traditional deconvolution process to overcome the staircase in the generated data. The TTGM was trained on non-void SRTM (Shuttle Radar Topography Mission) 1-arc-second data patches in mountainous regions collected across the globe. Then, information neighboring the voids was involved in order to infer the latent encoding for the missing areas approximated to the distribution of training data. This was implemented with a loss function composed of pixel-wise, contextual, and perceptual constraints during the reconstruction process. The most appropriate fill surface generated by the TTGM was then employed to fill the voids, and Poisson blending was performed as a postprocessing step. Two models with different input sizes (64 × 64 and 128 × 128 pixels) were trained, so the proposed method can effciently adapt to different sizes of voids. The experimental results indicate that the proposed method can obtain results with good visual perception and reconstruction accuracy, and is superior to classical interpolation methods.