Spatiotemporal Prediction of Vehicular Emissions Using Advanced Deep Learning Models on Large Road Network

As global economic growth and urban development continue, vehicle-derived air pollution is expected to escalate, despite technological advancements improving vehicle emissions performance. The increasing number of vehicles poses a significant threat to urban air quality, making the prediction of motor vehicle emissions based on road traffic and climate change crucial. This study aims to predict hourly spatiotemporal emissions for each vehicle type on every road segment across North Carolina, using deep learning models with a spatial resolution of 1 km² and an hourly temporal resolution. The meteorological and traffic-related data were extracted from the DANA tool and Google Earth Engine for June 2019. The methodology included selecting a time window from a year-long emissions dataset and identifying relevant features using correlation analysis. These features were then spatially joined in ArcGIS Pro and restructured to develop 25 model combinations for five pollutants—CO, NOx, CO2, PM2.5, and PM10—across five classified vehicle types. Each emission from each vehicle type was predicted using three models: Long Short-Term Memory (LSTM), Modified Residual Networks with Convolutional Neural Networks (ResNet42-CNN), and a hybrid model combining ResNet42-CNN with LSTM (ResNet42-CNN-LSTM). The models were evaluated using RMSE and R-squared metrics. The findings show that among all evaluated models, the ResNet42-CNN-LSTM model delivers the best performance, particularly for motorcycle NOx emissions, achieving an R-squared value of 77%. In addition to contributing to the field of deep learning for time series forecasting, this research introduces an innovative approach to predicting link-level emissions by vehicle type, demonstrating superior predictive accuracy, particularly for the complex four-dimensional nature of the data.