Machine Learning-Based Predictive Model of Ground Subsidence Risk Using Characteristics of Underground Pipelines in Urban Areas

In this study, a machine learning-based prediction model was developed using the attribute information of underground pipelines and the history information of ground subsidence in order to predict the risk level of ground subsidence in urban areas. The target area was divided into a grid with sizes of 100m
$\times 100\text{m}$
, 300m
$\times 300\text{m}$
, and 500m
$\times 500\text{m}$
, and the attribute information of underground pipelines in the grid and ground subsidence data were utilized to build a dataset. For input data, the pipeline’s diameter, the number of years used, and density were selected based on the pipeline’s length as the basic unit. Additionally, the risk level of ground subsidence was determined as the output data using historical information. A total of 36 datasets were built according to the conditions, and factors with significant correlation were selected through a correlation analysis of the datasets. The developed datasets were divided into training data and evaluation data. The synthetic minority oversampling technique was used to resolve the data imbalance. The model performance evaluation indexes used in this study were F1-score and AUC(Area Under the Curve). The performance of each model was compared, and the comparison results showed that a model that applied a preprocessed dataset with 500m
$\times 500\text{m}$
grid size, 10 years in use, 100mm pipeline diameter, and 1–2 ground sinks in Level 1 risk range to the LGBM(Light Gradient Boosting Model) classifier derived the best evaluation indexes(F1-Score:0.750, AUC:0.840). The map was found to be effective for predicting the risk level of ground subsidence in urban areas.