Risk Assessment Of Land Subsidence-Induced Structural Damage Of Unreinforced Masonry Buildings Across The Dutch Urbanized Area (2019-2023).

Abstract

Land subsidence (LS), the gradual lowering of the Earth’s surface, poses major challenges in low-lying, peat- and clay-rich urban regions such as Noord-Holland and Zuid-Holland in the Netherlands. It results from natural processes like subsurface consolidation and peat oxidation, as well as human activities including groundwater extraction, drainage, and structural loading. LS leads to structural damage and infrastructure failure. Satellite-based monitoring, particularly InSAR, provides millimeter-level deformation measurements over time, enabling land subsidence risk assessment in urbanized areas.

This study develops a data-driven machine learning framework to assess LS-induced structural damage for pre-1970 unreinforced masonry buildings on shallow foundations in soft Holocene soils. A Random Forest regressor, effective at capturing non-linear relationships, was trained using environmental, climatic, subsurface, and building-related predictors—elevation, soil characteristics, groundwater levels, land use, and urban infrastructure—to predict mean subsidence rates (mm/yr) for 2019–2023.

The model performed strongly on national training data (R² = 0.78, RMSE = 0.46, MAE = 0.30) and moderately on national testing data (R² = 0.45, RMSE = 0.73, MAE = 0.49), reflecting local heterogeneity not fully captured by predictors and smoothing in the target variable. Feature importance analysis identified climatic and topographic variables as dominant, with temperature most influential (23.0), followed by elevation (AHN4; 14.0) and precipitation (13.5). Geological and landscape factors also contributed, including GLG (11.9), GLG–GHG difference (8.7), and Holocene thickness (HT; 7.6). Building characteristics, such as total area (6.0), and urban land-use features—vegetation cover (Green; 5.8), roads (4.7), and water bodies (4.6)—provided additional predictive power.

Predicted subsidence patterns reproduced broad spatial trends and highlighted neighborhoods with higher LS rates. However, extreme local values and differential subsidence were consistently underestimated due to model smoothing. Final risk maps were derived using Risk = Hazard × Exposure × Vulnerability, where hazard represents differential subsidence, exposure corresponds to pre-1970 buildings, and vulnerability is defined by their short dimension. The resulting maps quantify average damage probability at neighborhood and municipal scales.

This research demonstrates the potential of integrating 100 m grid-cell EGMS data with machine learning to estimate structural damage probability in pre-1970 buildings. Yet, translating observational data into actionable urban risk assessments remains challenging due to complex interactions among subsurface properties, groundwater dynamics, building density, and structural characteristics. Model performance and spatial accuracy could improve with higher-resolution data, including detailed temperature and precipitation, refined building attributes, and finer-scale groundwater and subsurface datasets. Expanding temporal coverage and updating land-use information would further enhance land subsidence risk management.