Machine learning driven by environmental covariates to estimate high-resolution PM2.5 in data-poor regions

Area of Science:

• Environmental Science
• Atmospheric Science
• Public Health

Background:

• Particulate Matter (PM2.5) affects air quality and public health.
• Spatial distribution of PM2.5 is poorly understood in data-scarce regions.
• Xinjiang faces challenges in PM2.5 monitoring due to limited stations.

Purpose of the Study:

• To estimate the spatial distribution of PM2.5 concentrations in Xinjiang from 2015-2020 at 1 km resolution.
• To compare the performance of Random Forest (RF) and bagging algorithm models for PM2.5 estimation.
• To identify spatial and seasonal patterns of PM2.5 in Xinjiang.

Main Methods:

• Developed and validated Random Forest (RF) and bagging algorithm models.
• Utilized ground-monitored PM2.5 data, aerosol optical depth (AOD), meteorological data, and geographical variables.
• Employed 10-fold cross-validation (CV) for model verification.

Main Results:

• The RF model demonstrated superior performance for high-resolution PM2.5 estimation.
• PM2.5 concentrations were higher in southern Xinjiang (Tarim Basin) and lower in northern Xinjiang.
• Significant seasonality was observed, with winter concentrations highest (71.95 µg m⁻³) and summer lowest (43.40 µg m⁻³).

Conclusions:

• The RF model effectively estimates PM2.5 in data-scarce regions.
• The study provides crucial insights into Xinjiang's air quality patterns.
• This approach supports air quality monitoring and sustainable development efforts.