Spatial deep learning for PM 2.5 estimation during extreme pollution events: Quantifying uncertainty and training with sparse datasets
Year: 2026
Type: article
Abstract: Accurately estimating surface PM _2.5 during extreme pollution days (PM _2.5 > 35 μ g m ^−3 ) with confidence remains a challenge due to the lack of uncertainty quantification and sparse ground monitoring. This study develops a spatially aware deep learning model that integrates overlapping feature maps and convolution-based architectures to improve PM _2.5 estimation, quantify its uncertainty, and simulate sparse observation scenarios. Using gap-free MERRA-2 reanalysis data for the contiguous United States (2012–2023), we compared three model classes: (1) a pointwise artificial neural network (ANN) with no spatial context; (2) a vectorized ANN leveraging spatial information as model inputs; and (3) convolutional neural networks (CNNs) that learn spatial dependencies, with residual and attention blocks. By comparing the average model performance and its spatial and seasonal distribution, the residual CNN model using spatial context ∼400 × 400 km ^2 outperforms all other models, highlighting that encoding spatial context within the model structure is more significant than simply providing spatial information as inputs. When further quantifying the uncertainties of PM _2.5 through the 95% confidence interval (CI), the CIs’ interquartile ranges and medians for models considering spatial information ∼400 km are reduced by half compared with those with ∼200 km, likely due to the increasing size of the overlapping feature maps, from up to 16–64 estimations at each location. Lastly, the study evaluates the model robustness by removing locations without ground monitors (∼80% of the original data) and further by randomly removing an additional 50%–75% spatially. Even when trained on only ∼5% of the original data, the residual CNN maintained strong performance with R ^2 = 0.83 at held-out sites. Overall, this study establishes a spatial deep learning for PM _2.5 estimation with uncertainty quantification and sparse simulation, providing confidence for future integration with sub-daily satellite and ground observations to support air-quality monitoring and public-health applications.
Source: Machine Learning Earth
Authors: Shijin Wei, Kyle Shores, Yu-Sheng Chen, Yangyang Xu
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: gold
Estimated APC paid: $3,125
Award: G6094304093
OpenAlex ID: W7125599124