Diff-DEM: A Diffusion Probabilistic Approach to Digital Elevation Model Void Filling

Kyle Shih-Huang Lo

doi:10.1109/LGRS.2024.3403835

Diff-DEM: A Diffusion Probabilistic Approach to Digital Elevation Model Void Filling

2024-01-01

1

PRE

AI

DOI：10.1109/LGRS.2024.3403835

Original

Original request for help

Save

Abstract

En 中文

Digital elevation models (DEMs) are crucial for modeling and analyzing terrestrial environments, but voids in DEMs can compromise their downstream use. Diff-DEM is a self-supervised method for filling DEM voids that leverages a Denoising Diffusion Probabilistic Model (DDPM). Conditioned on a void-containing DEM, the DDPM acts as a transition kernel in the diffusion reversal, progressively reconstructing a sharp and accurate DEM. Both qualitative and quantitative assessments demonstrate that Diff-DEM outperforms existing DEM inpainting, including generative adversarial network (GAN) methods, inverse distance weighting (IDW), Kriging, LR B-spline, and Perona-Malik diffusion. The comparison is on Gavriil's and on our benchmark that expands Gavriil's dataset from 63 to 217 full-size ( 5051 x 5051 ) 10-m GeoTIFF images sourced from the Norwegian Mapping Authority; and from 50 DEMs to three groups of 1 k each of increasing void size. The code and dataset are available at https://github.com/kylelo/Diff-DEM.

Keywords:

Training

Kernel

Benchmark testing

Splines (mathematics)

Filling

Generative adversarial networks

Noise reduction

Denoising diffusion probabilistic model (DDPM)

digital elevation model (DEM)

generative model

inpainting

Diff-DEM: A Diffusion Probabilistic Approach to Digital Elevation Model Void Filling

Abstract

Journal

Organization