National Land Use Regression Model for NO₂ Using Street View Imagery and Satellite Observations.

Area of Science:

• Environmental Science
• Geospatial Analysis
• Machine Learning

Background:

• Land use regression (LUR) models are standard for estimating urban air pollution.
• Traditional LUR models use neighborhood-scale data from curated geodatabases.
• There is a need for more granular and scalable air quality modeling approaches.

Purpose of the Study:

• To develop national NO₂ models using street-level predictors from Google Street View (GSV) imagery.
• To compare the performance of GSV-only models with models incorporating satellite data (OMI).
• To evaluate the potential of GSV imagery for capturing intra-urban NO₂ variation.

Main Methods:

• Developed national NO₂ models using machine learning (random forest).
• Utilized street-level features extracted from GSV imagery as primary predictors.
• Created two model types: GSV-only and GSV + OMI (satellite NO₂ observations).
• Employed random 10-fold cross-validation to assess model performance (R²).

Main Results:

• GSV-only models achieved a cross-validation R² of 0.88, indicating street view data alone is highly informative.
• GSV + OMI models reached an R² of 0.91, showing satellite data offers incremental improvement.
• Model performance is comparable to traditional LUR approaches.
• Street-level features demonstrated potential for better intra-urban NO₂ variation capture.

Conclusions:

• Street view image-based modeling offers a powerful, unified framework for large-scale air quality modeling.
• GSV imagery alone can explain a significant portion of NO₂ variation.
• A cost-effective image sampling strategy is proposed for future studies.