Atmospheric pollution causes a thin/thick layer of dust/smoke (Haze) and hence obscures the visibility of the sky. This causes many issues for the countries varying from and including, but not limited to, imposing hazardous warnings from governments, closure of schools, and causing lung cancer. While there are many local causes of haze and fog, the impact of Transboundary hazes is often neglected, especially in the cases where the gases of forest fires or burning in one region expands and moves across the boundaries. This indicates the need to study Haze and Transboundary Haze in detail to understand their causes and develop a system based on insights to predict Haze in advance. Therefore, this paper is aimed to first explore the Haze data, containing PM-10 values, along with weather variables, and investigate the Time series methodologies to predict PM10 (Haze pollutant) values, given the same countries data and the data from other neighbouring countries(Transboundary Haze) as well. Further, methodologies including Artificial Neural Network (ANN) and Transfer Learning were employed to better generalize the models, learn from different experiences and apply the experience where data is scarcely available. The implementation and models based on different Time Series methods produced the SMAPE scores of 44.96, 29.07 and 27.74 on the Brunei, Singapore and Thailand data, respectively. However, with the use of cross-data, hyper-parameter tuning and Transfer learning, the results were improved to the best SMAPE scores of 25.08, 25.27, and 27.74 on the respective three datasets. These results were further presented in the MediaEval 2021 challenge and earned the Distinctive award for the detailed Pre-processing and validation steps, and top ranking in the task challenge.
View source
