Impact of grid size on spatiotemporal prediction of fine particulate matter.
Incorporating high spatiotemporal resolution in prediction models can significantly improve the accuracy of results. The current study focuses on identifying the most suitable spatial resolution using hexagonal tessellations of different sizes over the study area. Hexagonal grids are selected after comparing different types of grids. Spatial interpolation methods, including inverse distance weighting and ordinary Kriging method, are used for creating augmented datasets for each grid. A hybrid deep-learning-based prediction model is developed combining Convolutional Neural Network (CNN) and Long Short-term Memory (LSTM) for multivariate time series regression to predict PM2.5. Three scenarios are considered based on the size of the hexagonal grids. The results from each model are compared using evaluation indices of Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The values range from 0.8 to 0.95, which shows that the hybrid model gives accurate predictions. Analysis of models for different-sized grids shows that for 1, 3, and 5 km grid sizes, 17%, 24%, and 6% of the cells have RMSE values less than 20, respectively. Similarly, 39%, 50%, and 46% of the cells have MAE values less than 16 for 1, 3, and 5 km grid sizes, respectively. The comparison showed that the 3 km grid benefits data aggregation per prediction accuracy and computational time. The predictive capability of models is not dependent on the spatial relationship between data from different static monitors. In the future, mobile monitoring data can be included to increase the data points and models’ accuracy.
