A comparative study of stochastic nature of anthropogenic aerosol using SARIMA and LSTM methods

The troposphere is a dynamic region characterized by intricate interactions among natural and anthropogenic processes.Among these, aerosol loading, particularly fine particulate matter (PM 2.5 ), poses significant challenges in understanding the greenhouse effect and complicates air quality assessment due to its spatiotemporal variability.This study presents a comparative stochastic modelling framework for forecasting anthropogenic aerosol concentrations using both parametric and non-parametric approaches over Alandur (13.0°N, 80.10° E), Tamil Nadu, India.Specifically, a Seasonal Autoregressive Integrated Moving Average (SARIMA) model and a Long Short-Term Memory (LSTM) neural network, were employed to forecast PM 2.5 levels using observational data from 2023 and 2024.SARIMA, a classical statistical model, effectively captured regular patterns and seasonal trends.In contrast, the LSTM model, a deep learning approach, demonstrated enhanced performance in modelling non-linear dependencies and temporal dynamics.Model evaluations based on Root Mean Square Error (RMSE), indicated that LSTM consistently outperformed SARIMA, underlining its effectiveness in capturing the stochastic and complex behaviour of aerosol concentrations.These findings underscore the potential of integrating advanced machine learning techniques with traditional time series models to enhance the accuracy of air quality forecasting and environmental risk mitigation strategies.