Diffusion models generate high-resolution elevation maps from low-resolution inputs guided by optical imagery

The method enhances 5-meter coarse digital surface models to 0.5-meter resolution by transferring structural details visible only in high-resolution spectral images.

PaperIndustrycs.CVarXiv:2609.11886v1 ↗

Armand Mihai Nicolicioiu · Dominik Narnhofer · Nando Metzger · Daniel Panangian · Ksenia Bittner · Konrad Schindler

ETH Zürich · German Aerospace Center (DLR)

Research Digest·13 September 2026·2 min read

Read the paper →

The authors propose a guided super-resolution approach for digital surface models (DSMs) using denoising diffusion, improving coarse 5 m DSMs to 0.5 m resolution by leveraging high-resolution optical imagery. Experiments on several Central European cities show the method produces DSMs with crisper building outlines and more detailed roof structures than conventional interpolation or filtering.

What they did

The authors developed a diffusion-based method to super-resolve coarse DSMs (5 m resolution) to fine (0.5 m) using high-resolution optical images as guidance. The diffusion model transfers information such as building edges and roof shapes from the spectral images into the elevation domain. They evaluated on urban areas in Central Europe.

Key findings

The super-resolved DSMs exhibit improved structural detail, including crisp building outlines and detailed roof structures, compared to baseline interpolation and filtering techniques.
The method accurately reconstructs surface geometry, as measured by quantitative metrics.
It effectively leverages foundational image priors from diffusion models to guide elevation reconstruction, outperforming conventional approaches.

Why it matters

High-resolution DSMs are crucial for urban analysis, 3D reconstruction, and infrastructure monitoring but are expensive to acquire. This work shows that combining widely available coarse DSMs and high-resolution optical images with diffusion models can produce detailed elevation data at a fraction of the cost, potentially democratizing access to high-resolution surface models.

Caveats

The evaluation is limited to Central European cities; generalization to other terrains or regions with different urban morphologies remains untested. The method relies on the availability of co-registered high-resolution optical imagery, which may not always be present. Quantitative comparisons against a broader set of state-of-the-art super-resolution methods are not detailed in the abstract.

Analysis

This work continues a trend of leveraging generative image models for geospatial data, similar to applications in satellite image super-resolution and terrain reconstruction. By using diffusion models, the method can hallucinate plausible details consistent with the optical guidance, going beyond simple interpolation. An open question is how well the model handles areas where optical imagery is obscured or has different spectral characteristics.