Automated Road Surface Classification in OpenStreetMap Using MaskCNN and Aerial Imagery

The classification of road surfaces using OpenStreetMap (OSM) is important for a range of practical applications, such as navigation, urban planning, and infrastructure analysis. However, road surface tags in OSM are frequently missing or incorrect, requiring automated approaches for wide-spread verification and classification. In this work, a machine learning methodology, using the National Agriculture Imagery Program (NAIP) aerial imagery is introduced for road surface classification based on categories such as asphalt, concrete, gravel, and dirt. We used a MaskCNN-based model to include segmentation masks to boost the feature extraction and enhance the classification accuracy. This includes pre-processing of data, color analysis of image content, hierarchical loss functions, and calibration of models for robustness towards different image quality and environmental changes. The training and evaluation process was performed via PyTorch Lightning, where accuracy, precision, recall, and confusion matrices were used to evaluate the models. The results show that the proposed method significantly improves the classification accuracy of road surfaces compared with traditional classification methods. The model proposed here achieved an accuracy of 92.3% with a performance better than that of the SVM Classifier (81.2%), Random Forest (83.7%), and U-Net (89.6%) (from other works). The results hint at plausible use cases in automated OSM data validation, routing efficiency and traffic planning. Further work will involve implementing this model in Java OpenStreetMap editor and adapting it to different areas of the world.