# Mapping at-risk transportation infrastructure assets using statistical and machine learning methods

## TL;DR

This paper presents a method to identify vulnerable transportation infrastructure using machine learning and geographic data to improve asset management.

## Contribution

A novel method for mapping at-risk highway embankments and slopes using machine learning and geospatial data.

## Key findings

- Random forest outperformed other models with perfect AUC, F1, and Accuracy scores.
- Elevation, distance from streams, NDVI, and precipitation were the top factors influencing HWS failures.
- The method enables efficient identification of vulnerable infrastructure for targeted maintenance.

## Abstract

Geotechnical assets such as highway embankments and slopes (HWS) are critical to the integrity of transportation infrastructure. However, they are largely overlooked by Transportation Asset Management programs in the United States. The HWS are vulnerable to landslides induced by several factors including frequent occurrences of extreme rainfall events. Therefore, mapping vulnerable HWS and developing an inventory will significantly help with infrastructure asset management. To this end, this research adopted proven geographical information systems based on landslide susceptibility mapping methods typically applied to hillside slopes, and a method for mapping at-risk HWS assets was developed. Several supervised machine learning (ML) classification models were developed and evaluated to accurately map at-risk HWS in the study area of central Mississippi. Digital Elevation Models (DEMs) created from remote sensing data obtained from satellites, drone sensors, and terrestrial LiDAR were utilized to develop rasterized causative factors. The causative factors used included: Geotechnical and geomorphological attributes, such as slope, aspect, curvature, elevation, normalized difference vegetation index (NDVI), soil composition, and terrain from DEM; and hydrological factors, including precipitation, distance from the stream, groundwater depth, and topographic wetness index. Known locations of failed and not-failed HWS were selected and rasterized, and the pixels were extracted as ground truth data. The rasterized causative factors were utilized as independent features to train the classification ML models for predicting HWS failure susceptibility. Models were evaluated by developing confusion matrices and using probabilistic metrics such as area under curve (AUC) score, F-1score, and Accuracy scores. Random forest outperformed the other models (AUC, F1, and Accuracy scores of 1.0). Probability threshold tuning was performed on the random forest model, and susceptibility maps with different thresholds were evaluated. An optimal threshold of 0.75 was used to balance false negatives and false positives in the predicted results, ensuring more reliable identification of hazard-prone slopes. The trained RF model revealed that the elevation, distance from streams, the NDVI, and precipitation were the top four factors influencing HWS failures in this study. The method allows for easy identification of vulnerable HWS across vast geographic areas. This method helps in effective fund utilization by doing targeted interventions and preventative maintenance efforts. Transportation agencies can implement this methodology on HWS at any location to strategize geotechnical asset management efforts.