Spatiotemporal aerosol prediction model based on fusion of machine learning and spatial analysis

Abstract

Abstract This study examined long-term aerosol optical thickness (AOT) data from the Moderate Resolution Imaging Spectroradiometer (MODIS) to quantify aerosol conditions on the Korean Peninsula. Time-series machine learning (ML) techniques and spatial interpolation methods were used to predict future aerosol trends. This investigation utilized AOT data from Terra MODIS and meteorological data from Automatic Weather System (AWS) in eight selected cities in Korea (Gangneung, Seoul, Busan, Wonju, Naju, Jeonju, Jeju, and Baengyeong) to assess atmospheric aerosols from 2000 to 2021. A machine-learning-based AOT prediction model was developed to forecast future AOT using long-term observations. The accuracy analysis of the AOT prediction results revealed mean absolute error of 0.152 ± 0.15, mean squared error of 0.048 ± 0.016, bias of 0.002 ± 0.011, and root mean squared error of 0.216 ± 0.038, which are deemed satisfactory. By employing spatial interpolation, gridded AOT values within the observation area were generated based on the ML prediction results. This study effectively integrated the ML model with point-measured data and spatial interpolation for an extensive analysis of regional AOT across the Korean Peninsula. These findings have substantial implications for regional air pollution policies because they provide spatiotemporal AOT predictions.

Keywords

• Satellite
• Aerosol
• Machine learning
• Spatial analysis
• Remote sensing