Critical Electrical Infrastructure Segmentation in Arctic Conditions

Arka Singh, Sreeraman Rajan, Marzieh Amini, James R. Green, Kevin Dick

Abstract: Most critical electrical infrastructure monitoring research takes place in temperate climates; however, significant infrastructure is deployed in remote and harsh environments, such as within arctic climates. Automatic segmentation of power poles from vehicle-based imaging using deep learning methods promises to reduce manual effort and enable more frequent inspections. Previous studies have demonstrated the successful application of deep learning techniques for pole segmentation using UAV and ground-based imagery. The HRNet-OCR model, originally trained for scene segmentation of ground-based imagery for multiple objects, including poles, has demonstrated the adaptability to learn semantic segmentation of poles in Google Street View (GSV) images captured during temperate weather conditions. This research study aims to evaluate the utility of the HRNet-OCR model architecture for semantic segmentation of utility power poles from ground-based images captured in arctic weather conditions, specifically in the Iqaluit region of Nunavut, Canada. Our work demonstrates promising performance of the HRNet-OCR model for power pole segmentation, and further finetuning suggests the model's ability to learn to segment other electrical equipment, such as pole-mounted transformers and crossarms. The proposed approach presents a promising solution for automating critical infrastructure maintenance in remote and harsh environments.