Enhancement of digital elevation models using tree-based ensemble machine learning algorithms

Okolie, Chukwuma

Enhancement of digital elevation models using tree-based ensemble machine learning algorithms

Thesis / Dissertation

2023

Abstract

Geospatial practitioners and national mapping agencies (NMAs) in Africa are constrained to rely on open-access remote sensing datasets, even as they struggle to meet up with best practices on spatial data infrastructure and topographic map revision. Thus, global digital elevation models (DEMs) have gained worldwide prominence due to their free availability; a prime advantage when compared to prohibitively expensive airborne topographic surveys. However, the accuracies of global DEMs are affected by several anomalies that diminish their quality and compromise their adequacy for applications where precise and accurate terrain information is needed. This research proposes an explainable tree-based ensemble feature-level fusion framework for enhancing satellite DEMs using Cape Town, South Africa as a case study. The enhancement methodology combines elevation and terrain features data alignment (co-registration and resampling) with feature-level fusion (ensemble learning) into a DEM enhancement framework. The training datasets are comprised of eleven predictor variables including elevation, slope, aspect, surface roughness, topographic position index, terrain ruggedness index, terrain surface texture, vector ruggedness measure, percentage bare ground, urban footprints and percentage forest cover as an indicator of the overland forest distribution. The target variable (elevation error) was derived with respect to highly accurate airborne LiDAR. Initially, the qualities of two satellite-derived interferometric DEMs (NASADEM and Copernicus) and two photogrammetric DEMs (ASTER and AW3D) were comparatively examined in a series of qualitative and quantitative tests in five different landscapes spread across Cape Town: urban/industrial, agricultural, mountain, peninsula and grassland/shrubland. Based on their performances, Copernicus and AW3D DEMs were selected for further analysis. The next phase involved a comparative evaluation of ten treebased ensembles for enhancement of Copernicus DEM over agricultural lands. At two implementation sites, there was a 6 – 13% reduction in the MAE and 15 – 29% reduction in the RMSE, and the corrected Copernicus DEM showed several topographic improvements such as smoothing of rough edges, enhanced stream channel conditioning and diminution of coarse/grainy pixels. Following the comparison, three recent implementations of gradient boosting, the extreme gradient boosting (XGBoost), light boosting machine (LightGBM) and categorical boosting (CatBoost) were selected for the development of a robust DEM enhancement framework. After training and testing, the models were applied for correcting the DEMs at two implementation sites spread across the five landscapes. Going further, a rigorous hyperparameter tuning strategy was implemented for the three models using a principled, robust and computationally efficient Bayesian optimisation scheme. The optimisations were operationalised with ten steps of random exploration for diversification of the exploration space, and 40 - 50 iterations to increase the likelihood of finding an optimal combination of hyperparameter values. The uniqueness of the optimisation scheme is the very wide diversification of the search space for random exploration. The performance of the models was compared based on default hyperparameters versus Bayesianoptimised hyperparameters. The result is a sequential correction and fusion scheme to increase the vertical accuracy and reduce errors in the final DEMs. The corrections achieved significant and highly competitive accuracy gains of up to 64% RMSE (68% MAE) reduction in Copernicus DEM and up to 78% RMSE (82% MAE) reduction in AW3D DEM. The robustness of the proposed framework was proven in several performance evaluations and comparative assessments. Summarily, it outperformed a globally acclaimed corrected DEM and the authoritative South Africa national DEM, and surpassed the achievable accuracies of several previously proposed strategies, including multiple linear regression. Moreover, three-dimensional terrain analysis and lineament mapping showed the potential of the proposed scheme for enhancing deliverables in topographic and geologic mapping. The proposed approach also incorporates explainability measures to describe the interactions between predictor variables and their influence on the predicted DEM errors. It provides a cost-effective framework and ‘minimal' computation expense. The innovative DEM enhancement scheme proposed in this research is applicable to other global landscapes.

Keywords

Geomatics

Reference:

Collections

PhD / Doctoral

Full item page