Electrical Substation Detection from Satellite Imagery in the Netherlands

Mask R-CNN model achieving a tile F-score of 89.88% for detecting electrical substations from optical remote sensing imagery. The pipeline, developed by J J-F G DE PLAEN of Vrije Universiteit Amsterdam, extracts infrastructure data from OpenStreetMaps and trains models to support large-scale risk assessment studies. Experiments tested different under-sampling percentages and hyperparameter settings to address class imbalance.

Use Cases

• Training instance segmentation models for critical infrastructure detection based on satellite imagery and asset masks.
• Assessing climate-related risk hotspots for electrical substations based on geospatial detection data.
• Developing adaptation strategies for infrastructure resilience based on large-scale exposure mapping.
• Testing hyperparameter and class imbalance mitigation techniques for geospatial deep learning models.

Strengths

• Model achieved a tile F-score of 89.88% and an Average Precision at IoU 50% of 30.93.
• Pipeline methodology is described for processing imagery and asset masks from OpenStreetMaps.
• Experiments were conducted with specific under-sampling percentages (25%, 50%, 100%) for the majority class.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and dataset size are unknown, which may limit suitability assessment.
• Data may reflect geographic bias inherent to the case study focusing on the Netherlands.

Provenance

Source

Vrije Universiteit Amsterdam

Collection Method

Assets extracted from OpenStreetMaps, with imagery and masks processed to train a Mask R-CNN model.

Geography

The Netherlands