Hydrotechnical geohazards at pipeline watercourse crossings can compromise pipeline integrity, leading to leaks or ruptures. A conventional geohazard management strategy involves an initial field investigation of all identified watercourse crossings to collect information required to evaluate the hazard at each location, specifically the annualized probability of failure. Many of these locations require regular subsequent inspections to monitor changes over time.

In the last decade, the resolution and accuracy of mapped watercourses has greatly improved due to the increased availability of bare earth lidar and high-resolution satellite imagery. Creating an inventory of pipeline watercourse crossings now leads to the identification of a very large number of smaller watercourses, on average one crossing every three kilometres, making the cost of field inspections for all identified watercourse geohazards disproportionate to the overall risk posed. This paper explores a new approach using machine learning techniques to leverage widely available datasets to predict the probability of pipeline exposure and failure at small watercourse crossings, reducing the need for costly field inspections.

The machine learning model was trained on field inspection data collected over a ten-year period from over 20,000 pipeline watercourse crossings across North America. The model uses probabilistic algorithms to predict watercourse scour depths, pipeline burial depth, and pipeline vulnerability to failure, calibrated based on over three million pipeline kilometre-years of geohazard performance data. The model relies on widely available data that can be assembled at scale without conducting field assessments, such as hydrologic characteristics, surficial geology, land cover, and pipeline characteristics. The annualized probability of pipeline failure was calculated, allowing each site to be integrated into and prioritized within the broader pipeline integrity management program for Keyera Corp., a leading Canadian midstream gas pipeline operator.