Asset Ranking Using Drone-Based Inspection

This paper presents a novel approach to wooden utility pole maintenance prioritization using a specialized implementation of Siamese Neural Networks. Traditional pole inspection methods focus on the strength and condition of the pole at the ground-line. However most utility equipment is attached near the top of the pole, and attention to the condition of this area is lacking during inspections. While unmanned aerial vehicles (drones) enable automated pole-top inspection with computer vision, traditional object classification methods prove ineffective for wooden pole defect detection. Rather than attempting to fit defects into predefined categories, our methodology leverages human expert comparative judgments over absolute classifications, recognizing that pole defects vary greatly in severity based on size, depth, type, and location. This architecture, which we call Pairwise Encoder for Asset Ranking (PEAR), processes poletop images captured by drones, learning directly from pairwise preferences to generate normalized health scores. Trained on approximately 2,300 expert-labeled image pairs, our model achieves 86% accuracy for the pairwise comparison on hold-out data. The resulting score distribution reveals a bimodal pattern that aligns with observed pole deterioration behavior, where poles tend to remain in good condition before visibly deteriorating relatively rapidly once decay begins. This approach transforms subjective expert knowledge into an objective, scalable scoring system that prioritizes critical maintenance work without requiring comprehensive measurement data.