A Coastal European Dataset of Critical Infrastructure: Leveraging Deep Learning to Enhance Power Infrastructure Exposure Information for Disaster Risk Assessment.

Detailed information on the exposure of critical infrastructure (CI), such as power assets, is a necessity to establish accurate risk assessment from natural and human-made hazards. Currently, large-scale risk assessment mostly relies on Volunteered Geographic Information to establish the exposure of CI causing limited reliability due to inherent information gaps. Deep Learning offers the possibility to fill such gaps through the extraction of CI from remote sensing imagery.

Here we present a comprehensive high-resolution geospatial database encompassing key elements of the power grid, namely power towers, electrical substations, and power plants. The dataset is derived from a workflow using Worldview-2 0.4-meter resolution satellite imagery for the most populated urban areas along the European coastlines.

The method extracts infrastructure location from OpenStreetMap to create annotations. Subsequently, the satellite imagery raster and annotations undergo processing to constitute training data. Data augmentation is employed on the raster tiles to enhance the training dataset. The method then trains a Mask R-CNN model to automate the detection of CI. Additionally, saliency maps are generated to validate the proper functioning of the model.

Performance metrics, specifically mean Average Precision and F-scores of the tile classification, are presented to evaluate the model's ability to correctly identify and classify power infrastructure. Furthermore, to assess the completeness of the geospatial database, a comparative analysis is conducted with OpenStreetMap on “unseen” locations. This comparative study sheds light on potential gaps and discrepancies, offering insights into the overall reliability and comprehensiveness of the dataset.