Deep Learning-based Terrain Data Completion with Geomorphological Constraints

Raw terrain data acquired by sensing techniques such as SfM or LiDAR typically contain non-terrain components that require filtering, such as vegetation occlusion and other non-terrain features. While filtering helps remove non-terrain data, it can introduce discontinuities and local voids in the dataset. These data gaps can affect both the completeness of the terrain representation and subsequent analysis tasks. Therefore, it is crucial to develop effective terrain data completion methods for reliable terrain analysis.

Traditional terrain data completion methods, such as interpolation-based algorithms and Poisson surface reconstruction, typically model and optimize data continuity from a mathematical perspective. Although these methods address local voids to some extent, they generally fail to exploit terrain features and semantic information, limiting their effectiveness in completing complex terrain scenarios.

To address these issues, we propose a deep learning-based framework for terrain data completion. Our methodology explores different neural network designs with supervised and unsupervised learning, incorporating geomorphological constraints to improve terrain feature representation and semantic understanding. The framework leverages the representational capabilities of deep learning to improve the robustness of terrain data completion, contributing to a more consistent and reliable basis for subsequent terrain analysis and applications.