DSM2DTM: An End-to-End Deep Learning Aproach for Digital Terrain Model Generation

Remotely sensed Earth elevation data or digital surface model (DSM) typically contains both terrain and above-ground information such as vegetation and man-made constructions. However, many applications require pure bare-terrain data, also known as digital terrain model (DTM). But how do we separate 3D objects on the DSM from the ground? The most commonly used approaches are still based on various filtering techniques, which in turn involve the pre-definition of thresholds or specific parameters depending on the inhomogeneity of the scene. Despite many long existing and newly developed approaches the general fully automatic extraction of large-scale, reliable DTMs is still a problem – especially the preservation of steep terrain features in terraced landscapes. In this context, we explore several deep learning models and select one based on the EfficientNet architecture. This model serves as an encoder in the UNet-shaped framework and – despite its relatively low amount of parameters compared to common network architectures – it can automatically distinguish non-ground pixels and estimate the bare-ground height information while maintaining the complexity of the anthropogenic geomorphology of landscapes. In a series of experiments, we demonstrate that the DTM generated with the proposed method significantly outperforms other DTM generation approaches – both quantitatively and qualitatively