Saturday 15 March 2025
The quest for accurate digital surface models (DSMs) has long been a challenge in remote sensing and geospatial analysis. These 3D representations of Earth’s topography are crucial for urban planning, vegetation studies, and 3D reconstruction, but often suffer from voids or missing data due to occlusions, shadows, and low-signal areas.
To tackle this problem, researchers have developed various methods for filling these voids, ranging from traditional interpolation techniques to deep learning-based approaches. However, most of these methods focus on digital elevation models (DEMs) rather than DSMs, which present a more complex challenge due to the intricate structures present in DSMs.
In a recent paper, a team of researchers has introduced a novel approach for void filling in DSMs by adapting deep anisotropic diffusion models originally designed for super-resolution tasks. This method, dubbed Dfilled, leverages optical remote sensing images through edge-enhancing diffusion to inpaint DSMs.
The key innovation behind Dfilled is its ability to propagate contextual information into voids while preserving critical structural details such as sharp building edges and smooth transitions in natural terrain. To achieve this, the model uses a heat equation-based approach to refine the void initialization process and incorporates Perlin noise to generate realistic missing data scenarios.
The authors tested their method on both simulated and real DSM datasets, comparing it to traditional interpolation techniques and state-of-the-art deep learning methods. The results show that Dfilled outperforms these approaches in terms of accuracy and visual quality, effectively handling complex features and maintaining terrain integrity.
One of the most impressive aspects of Dfilled is its ability to adapt to different types of masks and void patterns, making it a versatile tool for various applications. Additionally, the method’s edge-enhancing diffusion mechanism helps to preserve fine-scale structural details, resulting in more realistic and accurate DSM reconstructions.
The implications of this research are significant, as accurate DSMs have far-reaching applications in fields such as urban planning, environmental monitoring, and disaster response. By providing a more effective means of filling voids in DSMs, Dfilled has the potential to revolutionize these areas and enable more precise and detailed analysis of Earth’s surface.
In practice, Dfilled can be used to fill voids in DSMs generated from stereo satellite imagery or other sources, allowing for more accurate representations of complex terrain features.
Cite this article: “Filling Voids in Digital Surface Models with Dfilled”, The Science Archive, 2025.
Digital Surface Models, Void Filling, Deep Learning, Anisotropic Diffusion, Optical Remote Sensing, Edge-Enhancing Diffusion, Perlin Noise, Heat Equation, Stereo Satellite Imagery, Geospatial Analysis
