A comparison of the parametric and non-parametric deep learning algorithm in estimating PM2.5 over Delhi-NCR region in north-west India

Abstract:

The national capital region (Delhi - NCR) of India is one of the highest polluted regions of the world with an annual average PM_2.5 concentration of 125 µgm^-3. With increasing population and industrialization, the problem of air pollution has compounded in recent years. In this study, the efficiency of the soft computing algorithm was examined using satellite derive aerosol optical depth (AOD) in conjunction with meteorological parameters and land use information to estimate the monthly mean PM_2.5 concentrations across the region from June 2017 to May 2018. The daily ground-level PM_2.5 data set was obtained from 21 different monitoring stations across Delhi-NCR from June 2017 to May 2018. A data-driven adaptive neuro-fuzzy soft computing algorithm was used along with the sub-clustering algorithm (ANFIS-SC) to estimate PM_2.5. For comparison purposes, PM_2.5 concentrations were also estimated using the artificial neural network (ANN) and classical multiple linear regression (MLR) method. Results showed a moderate to a high degree of agreement between observed and predicted PM_2.5 concentration for both testing and validation period (R² ~ 0.98 for ANFIS-SC, 0.85 for ANN and 0.64 for MLR). Comparative analysis showed that ANFIS-SC with the highest predictive ability and low fraction bias (-0.00004) compared to the ANN and MLR. The spatial modelling of the PM_2.5 concentration near the ground also revealed notable space-time variations. While comparing, the agreement of the estimated values with the in-situ observations shows a R² of 0.86 at an annual scale. Nevertheless, the model shows a R² of 0.74, 0.84, 0.70 and 0.76 for monsoon, post-monsoon, winter and summer, respectively.