Reliability and stability of a statistical model to predict ground-based PM 2.5 over 10 years in Karachi, Pakistan, using satellite observations

Understanding the complex mechanisms of climate change and its environmental consequences requires the collection and subsequent analysis of geospatial data from observations and numerical modeling. Multivariable linear regression and mixed-effects models were used to estimate daily surface fine particulate matter (PM 2.5 ) levels in the megacity of Pakistan. The main parameters for the multivariable linear regression model were the 10-km-resolution satellite aerosol optical depth (AOD) and daily averaged meteorological parameters from ground monitoring (temperature, dew point, relative humidity, wind speed, wind direction, and planetary boundary layer height). Ground-based PM 2.5 was measured in two stations in the city, Korangi (industrial/residential) and Tibet Center (commercial/residential). The initial linear regression model was modified using a stepwise selection procedure and adding interaction parameters. Finally, the modified model showed a strong correlation between the PM 2.5 –satellite AOD and other meteorological parameters ( R 2 = 0.88–0.92 and p -value = 10 −7 depending on the season and station). The mixed-effect technique improved the model performance by increasing the R 2 values to 0.99 and 0.93 for the Korangi and Tibet Center sites, respectively. Cross-validation methods were used to confirm the reliability of the model to predict PM 2.5 after 10 years.