Automated road surface classification in OpenStreetMap using MaskCNN and aerial imagery

Area of Science:

• Computer Science
• Geographic Information Science
• Remote Sensing

Background:

• OpenStreetMap (OSM) road surface data is crucial but often incomplete or inconsistent.
• Automated validation and classification of road surfaces are needed for various applications.
• High-resolution aerial imagery and deep learning offer potential solutions.

Purpose of the Study:

• To develop and evaluate a deep learning model for automated road surface classification.
• To improve the accuracy and completeness of OSM road surface data.
• To classify road surfaces into asphalt, concrete, gravel, and dirt types.

Main Methods:

• A MaskCNN-based deep learning model with attention mechanisms and hierarchical loss was proposed.
• National Agriculture Imagery Program (NAIP) aerial imagery was used with aligned OSM labels.
• Preprocessing involved georeferencing, data augmentation, label cleaning, and class balancing.

Main Results:

• The model achieved 92.3% overall accuracy and 83.7% mean Intersection over Union (mIoU).
• Performance surpassed baseline models (SVM, Random Forest, U-Net) and state-of-the-art methods.
• High precision and recall were obtained for all surface types, including gravel and dirt.

Conclusions:

• Combining NAIP imagery with attention-guided CNNs and hierarchical loss significantly enhances road surface classification.
• The model demonstrates robustness across diverse terrains and conditions.
• Potential applications include OSM data improvement, infrastructure analysis, and autonomous navigation.