Estimating fine particulate concentration using a combined approach of linear regression and artificial neural network

Fine particulate matter (PM2.5) is directly associated with the degradation of air quality and environmental health effects. PM2.5 is gaining much attention through its environmental impacts, but the inadequacy of ground based measurements limits the understanding of PM2.5 over many regions. This study is aimed to employ a new and integrated approach of multiple linear regression (MLR) and artificial neural networks (ANN) to estimate the ground level PM2.5 concentration using satellite aerosol optical depth (AOD), land use data and meteorological parameters. AOD from Moderate Resolution Imaging Spectroradiometer (MODIS) aerosol products (MOD04) with Dark Target Deep Blue Combined algorithm at 10 km spatial resolution were retrieved for the most urbanized and industrialized city of Karachi, Pakistan during 2015-2017. The results of the MLR model revealed a good agreement with the ground observed data through correlation (R) 0.96, 0.87 and 0.76 for 2015, 2016 and 2017, respectively. The ANN with error back propagation algorithm was developed using AOD with binning of land use and meteorological parameters with associated spatio-temporal terms. The data sets were assembled into three groups, with 80% data for training and 10% each for validation and testing. ANN revealed good correlation coefficients (R) 0.80, 0.80 and 0.78 for training, test and validation, respectively. The proposed study has shown the enhanced accuracy in estimating PM2.5 concentration by including meteorological and land use data with satellite AOD. The results showed that both MLR and ANN are in closed agreement and capable to estimate PM2.5 concentrations. Overall, for the estimation of particulate concentration, ANN is more powerful technique and can be used to estimate long term particulate matter concentration with associated guidelines to monitor air quality in any region.