Abstract

The Network Rail (NR) railway network is divided among 5 Regions and 14 Routes and covers 36,000 kilometers of track. There are more than 190,000 earthworks within the asset portfolio and over 40,000 are examined every year during manual inspections.

Network Rail’s 2014 national aerial survey project collected high resolution ortho-rectified imagery and a fully classified Light Detection and Ranging (LiDAR) point cloud. As LiDAR can penetrate through the small gaps in vegetation, it can return signals from the ground itself. The point cloud has been processed to remove all vegetation, buildings and other man-made structures to create a digital terrain model (DTM) of the ground surface. The DTM has an accuracy of 5cm vertically, therefore provides a valuable insight into earthworks that, from a photographic perspective, are often obscured by vegetation.

A counterfort drain is a drainage feature on a slope that lowers the groundwater within a slope and reduces the risk of the slope failing. They are often perpendicular to the alignment of the railway but can be diagonal. Counterfort drains are channels that are dug into a slope and are filled with aggregate to allow the controlled seepage of water down a slope towards a manage drainage system.

Earthworks examination, investigation and monitoring do not change the likelihood of failure of the earthworks, but they facilitate the risk-based approach and provide early warning of accelerating deterioration.

By better understanding their assets, National Rail can avoid getting to the point where massive remediations are required. A geospatially accurate and automated method for detecting counterfort drains would improve Network Rail’s asset inventory, enable asset engineers to make informed decisions, improve maintenance regimes and ensure compliance with the geospatial standard.

This challenge aimed to automatically detect counterfort drains in the provided aerial photography and digital terrain model (DTM) data. Whilst many drains are not visible in red-green-blue (RGB) images, they are identifiable in DTM data. The challenge was to build a machine learning model that can detect and segment drains using some combination of the two data modalities.

Citation information

Data Study Group team. (2022, April 27). Data Study Group Final Report: Network Rail. Zenodo. https://doi.org/10.5281/zenodo.6498538

Additional information

PIs: Rebecca Stone, Usman Nazir