Correcting Global Elevation Models for Canopy and Infrastructure Using a Residual U-Net

Abstract

Digital Surface Models (DSMs) are commonly employed to investigate topographical characteristics and processes; however, the presence of canopy and infrastructure in urban and forested areas can lead to height biases and inaccuracies. In this study, I aim to correct such biases by applying a deep learning approach known as Residual U-Net to remove the selected pixels and generate Digital Terrain Models (DTMs) that accurately represent the Earth's surface without canopy and infrastructure influence.

The Residual U-Net model was trained and tested on a dataset of DSM and DTM pairs, which were acquired from resampled AHN4. The model was evaluated on its ability to predict DTMs from DSMs, and its performance was compared with other existing methods. Additionally, the model was tested on different resolutions and the Copernicus DEM to assess its adaptability and generalization capabilities.

The results indicate that the Residual U-Net model outperforms conventional techniques, effectively reducing the influence of canopy and infrastructure, and resulting in DTMs with enhanced precision. The study also explores the errors in detail and identifies the model's error causes, highlighting its limitations and areas for potential improvement.

This study concludes by demonstrating the efficacy of applying deep learning techniques, such as Residual U-Net, to correct global elevation models for canopy and infrastructure. The results indicate that the model is a promising tool for topographical investigation in both urban and woodland situations, offering a versatile solution for generating accurate DTMs from DSMs.