Assessment of infrastructure deformation using EGMS-InSAR data and geo-environmental factors through machine learning: Railways and highways of Lombardy Region, Italy
Linear Infrastructures, characterized by high level of systemic vulnerability [1,2], are subject to several environmental and geological hazards. In the context of risk assessment and management, monitoring these important assets plays an important role in establishing the maintenance planning and preventive measures against the disruptive phenomenon, such as ground deformation due to natural and anthropogenic causes. In-situ and traditional infrastructure monitoring approaches, such as high-precision leveling measurements [3], are known to be costly and time-consuming. On the other hand, satellite Remote Sensing (RS) techniques, such as Synthetic Aperture Radar (SAR) Interferometry (InSAR), are recognized to be promising tools for monitoring and condition assessment of infrastructures [4].
As an essential branch of Copernicus Land Monitoring Service (CLMS), the new European Ground Motion Service (EGMS) is providing freely accessible ground deformation data spatially covering almost all European countries. The deformation time time-series contained in the datapoints are acquired based on InSAR processing of Sentinel-1 images from January 2016 up to December 2021 [5,6].
In this study, InSAR-derived deformation dataset, geo-environmental parameters, and Machine Learning (ML) techniques have been integrated to address the major causes of this complex phenomenon, specifically emphasizing railway and highway in Lombardy region, Italy. The vertical displacement velocities (mm/year) of EGMS datapoints located at the neighborhood of these infrastructures are utilized as the input ground motion data. The conditioning factors considered in this work include elevation, slope angle, slope aspect, precipitation, curvature, solar radiation, and Normalized Difference Vegetation Index (NDVI). The ML models, including Decision Tree (DT), Linear regression (LR), Light GBM (LG), XGBoost (XG), Random Forest (RF) and Extra Trees (ET), are used in this study. The Train-Test dataset ratio is considered to be 7:3, with respect to the higher performance of this ratio [7].
First, the used models have been validated using the Area Under ROC Curve (AUC), and ROC being Receiver Operating Characteristic curve. The results mostly show accep results (interval of 0.7 to 0.8) and the applicibility of the model. Then, the Relative Feature Importance (RFI) analysis is carried out to address the significant factors causing the ground deformatio. Also, the results regarding the Permutation-based and Shapley Additive Explanations (SHAP) importance decisions among the factors show that the rainfall (precipitation) and elevation are playing the most important role in the occurrence of the ground deformation detected on the infrastructures, based on the methodology adopted in this study. Also, the effect of solar radiation cannot be neglected. More detailed and further discussion of the results will be provided in the full version of this letter.
