Forecasting of aerosol optical depth over arid and semi-arid regions using time series and machine learning models.
Singh Sukhpal S, Parmar Kulwinder Singh KS, Soni Kirti K, Singh Sarbjit S
Aerosols play a crucial role in the Earth's radiation balance and represent a major source of uncertainty in global radiative forcing estimation. Accurate prediction of aerosol optical depth (AOD), a key parameter for characterizing atmospheric aerosols, is increasingly important due to rising air pollution and its environmental impacts, particularly in desert regions. This study investigates and compares advanced time-series forecasting models for AOD prediction across six major arid regions of Asia, including the Thar, Thal, Karakum, Kutch, Kyzylkum, and Arabian deserts. Considering the temporal variability of desert aerosols, three forecasting approaches were developed and evaluated: the classical statistical Autoregressive Integrated Moving Average (ARIMA), the deep learning-based Long Short-Term Memory (LSTM), and the dynamic Nonlinear Autoregressive with Exogenous Inputs (NARX) model. Initially, statistical analysis was performed to examine the temporal characteristics, variability, and distribution patterns of AOD series. The study used 294 monthly satellite-derived AOD550 observations retrieved using the Deep Blue algorithm, covering the period from February 2000 to July 2024. The performance of three forecasting models was evaluated using the 44-month test set, with root mean square error (RMSE) and the coefficient of determination (R2) used as performance metrics. Subsequently, 6-year-ahead AOD forecasts were generated for the period from August 2024 to July 2030. ARIMA showed capability in representing linear patterns but exhibited comparatively lower forecasting performance, with testing RMSE values ranging from 0.0250 to 0.0400 and testing R2 values between 0.40 and 0.60. LSTM achieved improved prediction accuracy over ARIMA by capturing nonlinear temporal dependencies, reducing testing RMSE to 0.0215-0.0350 and increasing R2 values to 0.53-0.70. The NARX model provided the best forecasting performance, achieving testing RMSE values between 0.0180 and 0.0275 and consistently outperforming ARIMA and LSTM across all desert sites. The Diebold-Mariano test confirmed that the superior performance of NARX was statistically significant. These findings demonstrate that nonlinear neural network-based models, particularly NARX, are more effective for capturing complex aerosol variability and provide a useful modeling framework for long-term AOD forecasting and environmental monitoring in arid regions. However, the moderate prediction accuracy of the models and the exclusion of additional meteorological variables highlight the scope for further improvement through the integration of relevant atmospheric parameters.