A data-augmentation approach to deriving long-term surface SO₂ across Northern China: Implications for interpretable...

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Data Science

Background:

• Northern China faced significant sulfur dioxide (SO2) pollution, hindering policy evaluation and health assessments due to data scarcity.
• Existing back-extrapolation methods are prone to bias, particularly

Purpose of the Study:

• To derive long-term, spatially resolved surface SO2 data for Northern China (2005-2019).
• To develop and validate a novel approach, robust back-extrapolation via data augmentation (RBE-DA), to address concept drift in SO2 modeling.
• To analyze the spatiotemporal trends and severity of SO2 pollution in the region.

Main Methods:

• Developed the robust back-extrapolation via data augmentation (RBE-DA) approach.
• Utilized satellite-retrieved SO2 column densities and employed interpretable machine learning for model diagnostics.
• Analyzed population-weighted SO2 ([SO2]pw) trends and spatial distributions from 2005 to 2019.

Main Results:

• Population-weighted SO2 ([SO2]pw) increased from 2005-2007, then decreased by 80.4% from 74.2 μg/m3 in 2007 to 14.6 μg/m3 in 2019.
• Severe SO2 pollution (>20 μg/m3) affected most of Northern China until 2017.
• Identified data imbalance in satellite SO2 column densities as a key factor causing estimation bias in traditional back-extrapolation.

Conclusions:

• The RBE-DA approach provides crucial long-term surface SO2 data for environmental policy and human exposure assessment in Northern China.
• Interpretable machine learning is vital for diagnosing and refining models, especially when dealing with data limitations.
• The RBE-DA method, leveraging satellite data, has global applicability for back-extrapolating various air quality measures.