Spatio-Temporal PM2.5 Forecasting in Thailand Using Encoder-Decoder Networks
Natch Sirisumpun, Kritchart Wongwailikhit, Pisut Painmanakul, Peerapon Vateekul
Abstract
PM2.5 is a type of particulate matter that contributes to air pollution in Thailand on a yearly cycle. Exposure to PM2.5 can cause acute health problems, including respiratory and cardiovascular diseases, as well as an increased risk of premature death. In this paper, we present a spatio-temporal model based on a deep learning approach for PM2.5 concentration prediction via an image-like approach at a country-wide level. Our model: SimVP-CFLL-ML is based on a video prediction model, called "SimVP". To enhance its performance when attempting to predict high PM2.5 concentration, SimVP includes two major improvements i.e. a cross-feature learning layer (CFLL) using 1x1 convolution layer to learn feature correlation and a masking layer (ML) to calculate loss in specific locations. The experiment is conducted on data collected from the pollution control department (PCD) of Thailand and sensor for all (SFA). Results show that our model outperforms all baselines. Our model’s F1 perforance is 3.51% better than the best baseline model for classifying high PM2.5 concentration class.