Enhancing the Vertical Accuracy of Copernicus Digital Elevation Model Using Tree-Based Machine Learning Models

EasyChair Preprint 9579, version 1

Abstract

The aim of this study is to assess the capability of tree-based supervised machine learning algorithms for enhancing the vertical accuracy of the 30m Copernicus DEM. Five models are compared with a 30m Copernicus-derived DEM in Cape Town, South Africa. The models used were Random forest (RF), AdaBoost (Adaptive boosting), XGBoost (Extreme gradient boosting), LightGBM (Light gradient boosting machine), and CatBoost (Categorical boosting). Initially, the Copernicus DEM was compared to a reference 2m aerial LiDAR dataset, and height residuals (∆h) were derived. Subsequently, the models were trained at five different sites using Copernicus-derived terrain parameters (elevation, slope, aspect, surface roughness, topographic position index, terrain ruggedness index, terrain surface texture, and vector ruggedness measure) and land cover parameters (urban footprints, percentage tree cover, and percentage bare ground cover) to predict the height residuals. At the test sites, analysis of the predicted residuals versus actual residuals showed a reasonable predictive capability of Random Forest, XGBoost, LightGBM and CatBoost, except for AdaBoost which performed below expectation at several sites. Thereafter, the trained models were applied for predicting height residuals and deriving corrected DEMs at five independent sites with similar terrain characteristics.

There is a remarkable reduction in the vertical errors achieved by the tree-based models. For example, the MAE of the original DEM in the grassland/shrubland area was reduced by 53% with LightGBM. Similarly, the RMSE and MAE in the urban/industrial area dropped by 42% and 45% respectively using Random Forest. The MAE reduction by XGBoost was 42% (urban/industrial) and 15% (cultivated fields). This shows the capability of tree-based ML models for improving the vertical accuracy of coarse resolution DEMs, with further gains in topographic mapping.

Keyphrases: AdaBoost, CatBoost, Copernicus, Digital Elevation Model, LightGBM, Random Forest, XGBoost

@booklet{EasyChair:9579,
  author    = {Chukwuma Okolie and Jon Mills and Adedayo Adeleke and Julian Smit},
  title     = {Enhancing the Vertical Accuracy of Copernicus Digital Elevation Model Using Tree-Based Machine Learning Models},
  howpublished = {EasyChair Preprint 9579},
  year      = {EasyChair, 2023}}