Learning Terrain Suitability for Rail using Portugal’s Existing Network

Railway alignment has its challenges, which is why there has been so much effort put into its automatization. These processes usually include a comprehensive geographic model, with a digital terrain model as one of its many layers.

This paper presents a technique to process the terrain data in a way that takes the directional component of its suitability into account. A Random Forest Model was trained using the existent rail network of Portugal as a mask for positive samples within a Digital Elevation Model and its directional derivatives. This allowed the construction of 8 directional cost grids, which then made it possible to apply a Dijkstra algorithm to test the model with possible paths.

The tests were made in Western Portugal, more specifically the connection between Porto and Lisbon, with the results showing clear signs of the model understanding real terrain suitability.