Point-surface fusion of station measurements and satellite observations for mapping PM2.5 distribution in China: Methods and assessment

Tongwen Li, Huanfeng Shen, Chao Zeng, Qiangqiang Yuan, Liangpei ZhangAtmospheric 2017 · DOI · arXiv

TL;DR

This paper introduces a national-scale mapping approach for $PM_{2.5}$ in China by fusing sparse ground station measurements with high-resolution satellite data. By leveraging a Generalized Regression Neural Network (GRNN) and a novel pixel-based merging strategy, the researchers achieved significantly higher estimation accuracy than traditional linear models, providing a clearer picture of air quality across China's diverse terrain.

What Problem Does This Paper Address?

While ground-level monitoring stations provide highly accurate $PM_{2.5}$ data, they are "point" measurements—sparse and unevenly distributed. To understand the health impacts of air pollution on a national scale, we need "surface" data: spatially continuous maps.

Distribution of ground PM2.5 monitoring sites across China.

Satellite-based observations of Aerosol Optical Depth (AOD) offer wide coverage, but the relationship between AOD (which measures particles in the entire atmospheric column) and ground-level $PM_{2.5}$ is notoriously complex. It is influenced by local meteorology, seasonal changes, and regional emissions. Previous studies often focused on small regions or used simple linear models that fail to capture these non-linear, large-scale variations. The authors set out to find a model that works across the entirety of China.

Methods

The core of the study is "point-surface fusion." The authors compared six different modeling approaches to link ground measurements with satellite data:

  1. Conventional Models: Corrected Linear Regression (CLR), Multiple Linear Regression (MLR), and a Semi-Empirical Model (SEM).
  2. Advanced Statistical/ML Models: Geographically Weighted Regression (GWR), Back-Propagation Neural Networks (BPNN), and the proposed Generalized Regression Neural Network (GRNN).

The GRNN is the standout feature. Unlike standard BPNNs, which can be slow to converge and get stuck in local minima, the GRNN is a type of radial basis function network that excels at function approximation with fewer parameters to tune.

To create the final maps, the authors proposed a pixel-based merging strategy. Satellite data often has "holes" due to cloud cover or high surface reflectance. This strategy uses interpolated ground data to maintain temporal trends while filling in these gaps, ensuring the final yearly and seasonal maps are spatially complete.

Datasets

The study utilized a multi-source dataset spanning February 2013 to December 2014, integrating ground-based monitoring with satellite and meteorological inputs.

Summary of datasets used in the study, including frequency and source.
  • Ground $PM_{2.5}$: Data from ~1500 monitoring sites across mainland China (CNEMC).
  • Satellite AOD: MODIS (Terra and Aqua) Collection 6 products at a high 3-km resolution.
  • Meteorology: MERRA-2 reanalysis data, including relative humidity, temperature, wind speed, surface pressure, and planetary boundary layer height.
  • Spatiotemporal Predictors: Latitude, longitude, and month were included to help the models account for regional and seasonal variations.

Key Insights and Ideas

  • Linearity is a Trap: Conventional linear models (CLR, MLR) performed poorly at the national scale. This suggests that the "AOD-$PM_{2.5}$ relationship" is not a single global rule but a complex, non-linear surface that changes across space and time.
  • The Power of Neural Networks: Machine learning models (BPNN and GRNN) significantly outperformed statistical models. They are better at "learning" the hidden influences of weather and geography on how particles settle near the ground.
  • Spring is the Hardest Season: All models performed worst during the spring. The authors attribute this to the influence of dust particles, which have different optical properties than the anthropogenic pollution the models are better at predicting.
  • Data Density Matters: The models were most accurate in Eastern China, where monitoring stations are dense. In the Northwest, where stations are sparse, accuracy dropped, highlighting the continued need for ground-level infrastructure.

Key Results

The GRNN model emerged as the clear winner for national-scale estimation:

  • Accuracy: In 10-fold cross-validation, the GRNN achieved an R-value of 0.816 and a Root Mean Square Error (RMSE) of 20.93 $\mu g/m^3$.
  • Comparison: Conventional models were much less reliable, with R-values hovering between 0.488 and 0.552.
  • Mapping Success: The final maps revealed a strong north-to-south decreasing gradient in pollution. The North China Plain was identified as the most polluted region (85–120 $\mu g/m^3$ annually), while Tibet and Hainan remained the cleanest (<35 $\mu g/m^3$).

Limitations and Future Work

Despite the improvements, the authors acknowledge several hurdles:

  • The "Cloud" Problem: Satellites cannot see through clouds. While the merging strategy helps, more advanced statistical methods for "gap-filling" missing AOD data are needed.
  • Missing Variables: Future iterations could improve accuracy by incorporating land-use data (e.g., proximity to roads or factories) and population density.
  • Long-term Trends: This study focused on a two-year window. The authors aim to extend this analysis over longer periods to support large-scale epidemiological studies on the long-term health effects of $PM_{2.5}$ in China.

Summary updated

BibTeX
@inproceedings{li2017point,
  author = {Tongwen Li and Huanfeng Shen and Chao Zeng and Qiangqiang Yuan and Liangpei Zhang},
  title = {{Point-surface fusion of station measurements and satellite observations for mapping PM2.5 distribution in China: Methods and assessment}},
  booktitle = {Atmospheric},
  year = {2017},
  doi = {10.1016/j.atmosenv.2017.01.004},
}
Figure