Storm-Drain and Manhole Detection Using the RetinaNet Method.

Area of Science:

• Urban Hydrology
• Computer Vision
• Geographic Information Systems (GIS)

Background:

• Storm drains and manholes are critical for urban hydrological modeling and flood mitigation.
• Accurate mapping of these features is essential for effective urban drainage system management.
• Deep learning (DL) offers promising solutions for automated feature extraction in urban environments.

Discussion:

• This study evaluates the RetinaNet object detection model for identifying storm drains and manholes in street-level RGB images.
• The performance of RetinaNet, utilizing ResNet-50 and ResNet-101 backbones, was assessed using mobile mapping data from Campo Grande, Brazil.
• Results indicate superior detection accuracy with RetinaNet employing the ResNet-50 backbone compared to Faster R-CNN.

Key Insights:

• RetinaNet with ResNet-50 demonstrates high accuracy in detecting storm drains and manholes from mobile mapping imagery.
• The DL approach effectively identifies key urban drainage components, crucial for hydrological analysis.
• The study validates the suitability of DL for mapping urban infrastructure from readily available RGB images.

Outlook:

• The developed methodology and labeled dataset can advance research in urban flood prediction and drainage system optimization.
• Further research can explore other DL architectures and datasets for broader applicability.
• This work contributes to the development of smarter urban infrastructure management systems.