Digital Elevation Models (DEMs) represent the Earth’s surface through elevation values and are used in urban planning, disaster management, agriculture, navigation, defense, hydrology, infrastructure planning, and environmental analysis. Their resolution directly affects the quality of downstream terrain analysis.

High-resolution DEM data is expensive and time-consuming to obtain, especially in data-poor regions. My graduation project explored whether deep learning-based super-resolution methods could improve low-resolution DEM data and make detailed terrain information more accessible.

Project Focus

The project implemented and compared Super-Resolution GAN (SRGAN) and Enhanced Super-Resolution GAN (ESRGAN) architectures for DEM resolution enhancement. The main goal was to reconstruct higher-resolution terrain elevation data from lower-resolution inputs and compare the results against traditional interpolation-based approaches.

The study focused on DEM data from Turkiye and Austria. Turkiye data was collected from selected flatter regions because of the country’s diverse terrain structure, while Austria provided a more balanced terrain dataset. The project also examined whether models trained on one geographic region could generalize to another.

Method

The DEM data was primarily based on Shuttle Radar Topography Mission (SRTM) data with 1 arc-second resolution, approximately 30 meters. The data was normalized and downsampled using interpolation techniques to create low-resolution inputs for training.

Both SRGAN and ESRGAN were trained with generator and discriminator networks. The generator reconstructed high-resolution DEM outputs from low-resolution inputs, while the discriminator evaluated whether the generated DEMs were realistic.

The training process used several loss functions:

  • adversarial loss
  • reconstruction loss
  • total variation loss
  • SSIM loss
  • MS-SSIM loss

These losses were used together to improve both numerical accuracy and structural quality in the reconstructed terrain data.

Evaluation

The models were evaluated using common image and reconstruction metrics:

  • Peak Signal-to-Noise Ratio (PSNR)
  • Structural Similarity Index (SSIM)
  • Mean Squared Error (MSE)

The experiments compared SRGAN and ESRGAN on Turkiye and Austria datasets, and also tested transferability by evaluating Austria-trained models on Turkiye data.

Outcome

The results showed that both SRGAN and ESRGAN improved DEM resolution compared with traditional interpolation methods. ESRGAN performed better overall, producing higher PSNR and SSIM values and lower MSE values in the reported experiments.

The project also showed that transferability between regions is possible when terrain characteristics are similar, although performance can drop when moving between geographic areas. This suggests that broader and more diverse training datasets could improve generalization.

For me, the project was an early exploration of a theme I still care about: using machine learning to improve the quality of information that other systems and people build on.