Predicting building types using OpenStreetMap

Area of Science:

• Geographic Information Science
• Machine Learning
• Remote Sensing

Background:

• Accurate building data is crucial for urban planning, humanitarian aid, and navigation.
• OpenStreetMap (OSM) offers extensive building geometry but lacks detailed semantic attributes like building type.
• Manual data enrichment is time-consuming and costly.

Purpose of the Study:

• To develop an automated, supervised learning approach for classifying building types in OSM data.
• To enrich volunteered geographic information (VGI) with semantic building attributes without manual intervention.
• To demonstrate the model's accuracy and transferability across different geographic regions.

Main Methods:

• A supervised learning model was developed to classify buildings as residential or non-residential.
• The model utilized existing OSM tags and incorporated geometric/topological features (e.g., footprint size, road adjacency, proximity to parking).
• Training and testing were performed using ground truth data from Fairfax County (VA), Mecklenburg County (NC), and Boulder (CO).

Main Results:

• The proposed approach achieved high accuracy in classifying building types within the study areas.
• The trained model demonstrated high transferability, maintaining accuracy in regions without ground truth data.
• This method effectively addresses the sparsity of semantic building information in OSM.

Conclusions:

• Automated building type classification using supervised learning is feasible and accurate.
• The model's transferability allows for broad application in enriching global OSM data.
• This work provides a valuable tool for the OSM and data science communities to enhance VGI.