Benchmarking Deep Learning Super-resolution Techniques for Digital Elevation Models in Mountainous Regions

RWTH Publications (RWTH Aachen)

Abstract

Original Resolution: 25 m grid spacing Patch Extraction: 128 × 128 pixels tiles, extracted with a stride of 64 pixels to form the high-resolution set Downsampling: Each tile subsampled by factors of 2× and 4× to create low-resolution variants Noise Augmentation: Additive Gaussian noise (μ = 0, σ = 2.5) applied to low-resolution tiles to create noisy datasets Digital Elevation Models (DEMs) are 3D representations of a terrain's surface, encapsulating critical information like elevation, slope, and aspect.Decision support systems and downstream applications rely heavily on the accuracy and resolution of the underlying data.Recent advancements in deep learning-based superresolution present an opportunity to significantly enhance DEM quality 4. Results 3. Experimental Setup GAN-Based (SRGAN) Attention-based (Swin2SR) CNN-based (SRCNN)

Authors 3

1. Nazanin Bagherinejad Aachen Model-based Development in Computational Engineering

   RWTH Aachen University

   Affiliation as printed

   Model-based Development in Computational Engineering , RWTH Aachen University

2. Antara Dasgupta Aachen Institute of Hydraulic Engineering and Water Resources Management

   RWTH Aachen University

   Affiliation as printed

   . Institute of Hydraulic Engineering and Water Resources Management , RWTH Aachen University

3. Julia Kowalski Aachen Model-based Development in Computational Engineering

   RWTH Aachen University

   Affiliation as printed

   Model-based Development in Computational Engineering , RWTH Aachen University