A deep learning pipeline to power infrastructure detection in high-resolution satellite images

Power grids are essential infrastructure systems that form the backbone of modern economies. Ensuring reliable and up-to-date data on power infrastructure is crucial for effective risk management and resilience planning, especially in the face of natural hazards that threaten the stability of these systems. This study introduces an open-source pipeline leveraging deep learning to detect power infrastructure–specifically power towers and poles–from high-resolution satellite imagery. By applying a Faster R-CNN model with a ResNet-101 backbone on Worldview-3 imagery over Vietnam, the study achieved good performance metrics, including an AP50 of 60.6%, an F1-score of 74.7%, and an accuracy of 90.9%. Additionally, the study developed a comprehensive, open-access dataset of power infrastructure annotations, setting a foundation for future research in this field. This research highlights the potential of deep learning models in mapping power infrastructure on a large scale, contributing to risk assessment and improved power system modelling. The framework’s adaptability and proven effectiveness highlight its potential for application across various geographical contexts and in detecting other types of critical infrastructure.