A Deep-Pollutant-Spatial-Operator-Network (DPSON) for spatial estimation of PM2.5, PM10, O3 and NO2, case study at Delhi, India

Atmospheric pollutants affect human health, disrupt ecosystems, and impact the economy. Spatial prediction of atmospheric pollutants using data from ground monitoring stations (GMS) is vital for informed decision-making and sustainable ecosystem management. To estimate atmospheric pollutants, this study introduces the Deep-Pollutant-Spatial-Operator-Network (DPSON) framework that combines GMS data with multi-source spatial covariates in order to produce precise predictions at a 1 km × 1 km grid across Delhi (Indian capital city). Pollutant data from 40  Central Pollution Control Board, India (CPCB) monitored GMS locations (January 2021–December 2022) were used for this purpose. The PM_2.5 and PM₁₀ datasets are available at a 3-hour resolution, while O₃ and NO₂ at a 1-hour resolution.

Normalized static spatial covariates, such as population density, waterbody concentration, road-length concentration, green cover, and Land Use Land Cover (LULC), were included to improve the DPSON model’s accuracy. To improve the dataset's generalization in relation to spatial covariate variations, additional samples were generated using the Sequential Gaussian Simulation (SGS) algorithm, randomly simulating pollutant observations at 100 grid locations on a 1 km² spatial grid for each timestamp and pollutant species, based on the pollutant concentrations observed at 40 GMS locations. These SGS-generated and GMS-observed datasets were combined for developing the DPSON model.

A specially crafted reference Distance-Assisted Location Embedding (DALE) approach was utilized to provide accurate spatial scaling and embedding of the locations within the DPSON network. The approach utilizes cosine and sine transformations of latitude and longitude, combined with a sine transformation of the distance from a reference point, to create suitable spatial embeddings for the network. The model architecture comprises two parameterized networks: (1) the Branch Network and (2) the Trunk Network. The Branch Network is responsible for embedding the pollutant data observed by GMS along with the static spatial covariates of the corresponding locations and their DALE. The Trunk network uses the DALE of unsampled locations, their static spatial covariates to estimate the pollutant concentration at those locations. The DPSON network’s reconstruction error (i.e.: Trunk network output) on the CPCB locations were considered for checking the model capability. The DPSON model was eventually compared with other baseline models. The proposed DPSON model achieved the following performance metrics: for PM_2.5, RMSE of 31.91 µg/m³, MAE of 18.35 µg/m³, and R² of 0.88; for PM₁₀, RMSE of 49.95 µg/m³, MAE of 32.02 µg/m³, and R² of 0.87; for O₃, RMSE of 11.75 µg/m³, MAE of 7.27 µg/m³, and R² of 0.85; and for NO₂, RMSE of 12.05 µg/m³, MAE of 7.67 µg/m³, and R² of 0.88. The proposed DPSON model outperforms all the baseline models for each of the pollutants and is adept at managing various types of spatial covariates, accommodating complex GMS observation distributions, while also providing a computationally efficient framework for the spatial estimation of pollutants.

 

Acknowledgement: We gratefully acknowledge that this study was supported by the “Indo-German Joint Research Collaboration” grant (DST/INT/DAAD/P-23/2023 (G)) from the Department of Science and Technology (DST), Government of India and DAAD, Germany