An improved geographic pattern based residual neural network model for estimating PM2.5 concentrations

Accurate and continuous PM_2.5 data is essential for effective prevention of PM_2.5 pollution. Despite the achievements of deep learning methods in estimating PM_2.5 concentrations, existing neural network models have relied too much on the self-learning capability and have ignored geographic patterns of PM_2.5. Few have taken a geographic perspective when modeling PM_2.5, resulting in lower model interpretability. In this paper, rather than inputting spatiotemporal information directly into the networks, we propose an improved geographic pattern based residual neural network (IGeop-ResNet) for estimating PM_2.5 concentrations in the Beijing-Tianjin-Hebei region (BTH) of China considering spatial heterogeneity and spatial autocorrelation by introducing spatial eigenvector and attention mechanism, as well as the encoding and embedding methods for temporal categorical variables. A DEM-weighted loss function was introduced to enhance the spatial predictive ability, particularly in high-altitude regions. The results show that the IGeop-ResNet model achieves excellent spatial predictive abilities (R² of 0.925 in terms of station-based cross-validation) and offers a certain level of interpretability compared to the Ori-STResNet (ordinary directly inputs temporal and spatial information in the ResNet model) and the Geop-ResNet model (without the DEM-weighted loss function). Continuous maps derived from the IGeop-ResNet model suggest the PM_2.5 concentrations in the BTH region exhibited a downward trend from 2015 to 2018 and experienced a sharp drop in 2017. The results indicate that NO₂ is the Granger cause of PM_2.5, while the relationship between SO₂ and PM_2.5 is insignificant.