Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
InSAR time series analysis and machine learning for ground subsidence monitoring and susceptibility mapping in Midvaal, South Africa
The Midvaal region in Gauteng faces a significant challenge of ground subsidence, which poses significant risks, including sinkhole formation, disruptions to infrastructure, and threats to human settlements. Despite the severity of this issue, existing methodologies for monitoring and mapping ground...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English Eng |
| Published: |
School of Architecture, Planning and Geomatics
2025
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| Summary: | The Midvaal region in Gauteng faces a significant challenge of ground subsidence, which poses significant risks, including sinkhole formation, disruptions to infrastructure, and threats to human settlements. Despite the severity of this issue, existing methodologies for monitoring and mapping ground subsidence susceptibility are limited by inefficiencies, labour-intensive processes, and reliance on sparse data sources, leading to delayed and inaccurate risk assessments. This study addresses these challenges by integrating Interferometric Synthetic Aperture Radar (InSAR) time series analysis with Machine Learning (ML) techniques to monitor ground subsidence and develop ground subsidence susceptibility maps for Midvaal, South Africa. The key research objectives included evaluating the effects of various InSAR pre-processing and processing software on ground subsidence monitoring accuracy, examining the impact of Digital Elevation Model (DEM) vertical accuracy on ground subsidence monitoring accuracy, identifying the spatial and temporal (spatio-temporal) patterns of ground subsidence, and employing machine learning to create predictive models for ground subsidence susceptibility mapping. Furthermore, the study provides actionable recommendations for rural-urban planning and infrastructure management based on its findings. This study assessed various InSAR processing methods, including SNAP, ISCE, HyP3, StaMPS, and MintPy. ISCE PS-InSAR was the most accurate for tectonic deformation monitoring, with minimal errors and stable velocity estimates. In contrast, HyP3 SBAS-InSAR performed well in detecting localised deformations but had higher error margins. Additionally, A semi-automated preprocessing tool, StoSAP, was developed using SNAP and Python-3 to streamline workflows for PS and SBAS methods. DEM vertical accuracy effect was another critical factor examined. Among the AW3D30, COP30, SRTMGL1, and ASTER DEM evaluated, AW3D30 displayed the highest vertical and velocity modelling accuracy, particularly in less complex terrains, whereas ASTER performed poorly in areas with dense vegetation, urban infrastructure, and mining activities. These findings underscore the importance of high-quality DEMs in enhancing ground subsidence monitoring and susceptibility mapping. Spatio-temporal analyses revealed that ground subsidence and uplift patterns were strongly influenced by land use, hydrological factors, vegetation cover, soil moisture, precipitation, and ground water storage anomalies. Subsequently, machine learning algorithms including Random Forest, XGBoost, LightGBM, and CNN were employed to create ground subsidence susceptibility maps and classify the region into five risk zones: Very Low, Low, Moderate, High, and Very High. Random Forest (RF) achieved the highest predictive accuracy with a Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and coefficient of determination (R² score) of 3.621 mm/year, 5.039 mm/year, and 84.98% respectively, identifying factors such as NDVI, lithology, and hydrological parameters as primary contributors to ground subsidence through SHapley Additive exPlanations and Frequency Ratio analysis. Furthermore, Random Forest classified the region into five risk zones, namely, Very Low, Low, Moderate, High, and Very High with values 0.1%, 87.9%, 9.4%, 1.7%, and 0.9% respectively indicating that the largest area in the study region is characterised by “Low” ground subsidence susceptibility. This suggests that the majority of the ground in the study region does not pose a significant threat to the environment. However, the combined occurrence of “High” and “Very High” susceptibility covers at least 2.6% of the area, equivalent to approximately 107.28 square kilometres about the size of Midrand, Gauteng, South Africa (152.87 square kilometres). This highlights a potentially alarming impact, given the significant proportion of high-risk areas. This study advances the understanding and management of ground subsidence by integrating InSAR time series analysis, machine learning, and environmental modelling into a robust, scalable framework for real-time monitoring and risk mitigation. The proposed methodology not only enhances the detection and prediction of ground subsidence in the Midvaal region but also holds potential for broader geohazard applications. Future research should explore the integration of multi-sensor data, real-time environmental monitoring, and further model refinement to improve performance in complex terrain. |
|---|