Mapping Utility Poles in Aerial Orthoimages Using ATSS Deep Learning Method.

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Traditional utility pole mapping using street-level imagery is labor-intensive and inefficient for large areas.
• Aerial imagery presents a scalable alternative, but detecting poles from top-view images in urban settings is challenging.
• Existing remote sensing object detection methods like Faster R-CNN and RetinaNet have limitations in accurately mapping utility poles.

Discussion:

• This study introduces Adaptive Training Sample Selection (ATSS), a novel method for detecting utility poles in urban environments using aerial imagery.
• The performance of ATSS was compared against Faster R-CNN and RetinaNet using a large dataset of 99,473 high-resolution image patches (10 cm GSD).
• The impact of bounding box size on ATSS performance was evaluated, with optimal results observed for larger bounding box dimensions.

Key Insights:

• ATSS demonstrated superior accuracy, achieving an Average Precision at 50% Intersection over Union (AP50) of 0.913, outperforming Faster R-CNN (0.875) and RetinaNet (0.874).
• Larger bounding box sizes significantly improved ATSS performance, with AP50 increasing by 6.5% and AP75 by 23.1% when comparing 60x60 to 30x30 pixel boxes.
• All tested methods exhibited comparable computational costs, processing approximately 0.048 seconds per patch.

Outlook:

• ATSS shows significant potential for developing automated tools for utility pole detection and mapping in urban infrastructure management.
• Further research could explore ATSS in diverse remote sensing scenarios and with varying image resolutions.
• Optimizing bounding box selection strategies within ATSS could further enhance detection accuracy and efficiency.