loading page

Generating DTM from DSM Using a Conditional GAN in Coastal Urban Areas of Japan
  • Haruki Oshio ,
  • Keiichiro Yashima ,
  • Masashi Matsuoka
Haruki Oshio
Tokyo Institute of Technology

Corresponding Author:[email protected]

Author Profile
Keiichiro Yashima
Author Profile
Masashi Matsuoka
Author Profile

Abstract

The recent surge in floods requires developing a digital terrain model (DTM) with a high spatial resolution that is essential for risk assessment at the building level. To generate a DTM, it is necessary to remove nonground objects from the digital surface model (DSM) and interpolate the elevation of the removed area. However, automatically conducting this process requires input data other than the DSM and/or setting parameters suitable for the target region. Here, we used the pix2pix model that performs domain conversion using conditional adversarial generative networks to directly generate DTM from only DSM. DSM and DTM with a spatial resolution of 1 m for coastal urban areas of Japan were used to train and test the model. The test results showed that the root mean squared error (RMSE) of the generated DTM was 0.4 m for the entire dataset. The RMSE increased with the increasing nonground object height and nonground object coverage ratio in an image. However, the RMSE was approximately 1 m or less up to a nonground object height of 50 m and a coverage ratio of 0.7. This article demonstrates that pix2pix is effective in translating DSM into DTM in coastal urban areas.