Prediction of risks for fires along railroad tracks using machine learning – BurnML

In the wake of climate change, fires along railways in Germany are now occurring more frequently, with sometimes massive consequences for operations and affected infrastructure. In the absence of systematic investigations, little is known about the causes of such fires. Since a further increase of these hazards is to be expected, there is a need for sound knowledge on triggers and drivers for fires as well as methodical knowledge on prediction tools. 

With this in mind, adelphi and the project partner Hochschule für Nachhaltige Entwicklung Eberswalde (HNEE) are pursuing the following goals with the BurnML project:

  1. to build a dataset for training a model, including data on past embankment fires.
  2. the training of a machine learning model for the prediction of fire hazards
  3. the pilot application of the model with DB Netz AG. Here, the team incorporates the following data, among others: satellite data on fires and land use, weather data from the DWD and infrastructure data from DB Netz.

Central activities of the project teams are the spatial identification of past embankment fires by means of evaluation of free satellite data, the recombination and analysis of the data on possible triggering and contributing factors for slope fires, selection and training of an algorithm for the analysis and prediction of such fires, evaluation of the findings from this machine learning approach, testing of the prediction model with DB Netz AG; always accompanying: discussion of (interim) results with stakeholders and publications.

As a result, adelphi and HNEE will publish, i.a., an extensive open geo-dataset, a documented machine learning classifier (in Python), a technical article with insights into triggers and drivers, and a prediction tool for embankment fire risks. The results expand the knowledge of embankment fires and prevention options and enable infrastructure companies to take precautions more efficiently.

The project is supported by a grant from the mFund funding programme of the Federal Ministry for Digital and Transport (BMDV).