Chutinun Potavijit. A Hybrid deep learning model for forecasting PM2.5 concentrations in northern Thailand from satellite images. Master's Degree(Applied Statistics and Analytics). Chiang Mai University. Library. : Chiang Mai University, 2568.
A Hybrid deep learning model for forecasting PM2.5 concentrations in northern Thailand from satellite images
Abstract:
Air pollution is a major environmental issue that causes widespread impacts, particularly fine particulate matter smaller than 2.5 microns, known as PM2.5, which poses serious health risks to the public. Northern Thailand has been significantly impacted by PM2.5 pollution, particularly during the dry season, mainly because of a result of agricultural burning practices and wildfires. This study intends to create hybrid deep learning models, specifically CNN-ANN, CNN-DNN, and CNN-LSTM, that can forecast PM2.5 concentrations by utilizing satellite imagery that represents four important environmental variables: aerosol optical depth (AOD), temperature, precipitation, and ozone (O3). These variables are recognized as being closely associated with the occurrence of PM2.5. The traditional models, such as MLR, SARIMAX, and individual deep learning models, are compared and evaluated against the proposed hybrid deep learning models. According to the research, the fundamental models are outperformed by all hybrid models in terms of predictive accuracy. In addition to them, the CNN-LSTM model demonstrated the most outstanding performance, with a coefficient of determination (R²) of 0.9982, a mean absolute percentage error (MAPE) of 0.94%, a mean absolute error (MAE) of 0.2497 µg/m³, and a root mean squared error (RMSE) of 1.0212 µg/m³. This investigation underscores the efficacy of hybrid deep learning methodologies in conjunction with satellite imagery to encourage the development of public policy and improve community preparedness in regions without ground monitoring stations, where the severity of PM2.5 pollution fluctuates during different seasons and regions.