Estimate Particulate Matter PM2.5 Concentration Impact of Wildfires Using Machine Learning in Chiang Mai Province, Thailand

Abstract

Wildfires are one of the most prominent problems with wide-ranging impacts on terrestrial ecosystems around the world. The important factor that causes air pollution is that the main cause is burning in open areas and large forest areas. Thailand experiences PM2.5 concentrations that are increasing every year, during the winter and dry season from December to April. Most of the concentration is concentrated in the central and northern regions of Thailand, especially Chiang Mai Province. PM2.5 has an effect on the economy and is very dangerous to the health of residents. However, air quality monitoring is often measured with a limited surrounding station. The insufficient number of monitoring stations is the challenge, rendering the measurement of PM2.5 concentrations less reliable and incongruent with the actual environmental conditions. This research is directly aimed at assessing PM2.5 concentrations resulting from wildfires using Remote Sensing data.

The main contents of this thesis include:

Evaluate the performance of the developed models in estimating PM2.5 concentrations at on-site scales and different seasons on regional scales within Chiang Mai Province, Thailand, in 2023.

comparing and determining optimal models for estimating PM2.5 concentrations, as well as identifying the primary factors influencing variations in pollution levels in regions impacted by severe wildfires.

Create a map of the spatial distribution of PM2.5 concentrations in areas that do not have ground measurement stations with remote sensing data using machine learning.

The results show that the Random Forest (RF) model demonstrates higher performance than the XGBoost and CNN models in estimating PM2.5 concentrations at on-site scale measuring, as evidenced by determination coefficients (R2) values of 0.74–0.91, RMSE values of 10.40–30.53 μg/m,3 and MAPE values 18.56–36.48 μg/m3, respectively. And the model demonstrated an average concentration all stations with an R2 value of 0.89, an RMSE of 11.61 μg/m³, and a MAPE of 34.22 μg/m³. Moreover, the RF model estimated the significance of features importance on PM2.5 concentration, including aerosol AOD-MAICA (MCD19A2) at 40%, AOT550 nm from MERRA-2 at 22%, dust mass PM2.5 from MERRA-2 at 12%, and CO from Sentinel-5P TROPOMI at 11%, ranking highest among all chemical components due to origin from combustion, aligning with the hypothesis that wildfires and greenhouse gas emissions significantly impact air quality. In addition, the RF model was used to predict values and create spatial distribution maps for Chiang Mai Province, Thailand, in 2023. The model's average accuracy had an R² of 0.81, RMSE of 14.45 μg/m3, and MAPE of 21.25 μg/m3.

The spatial distribution of PM2.5 concentrations reveals distinct patterns, with elevated levels observed in both the north zone and south zone of the study area. This is most consistent with data on wildfire activity. Heat maps of severe wildfires based on ground truth data show the actual heat points from combustion consistent with the same period of the month where the concentration of PM2.5 was high from January to April. This is the clearest confirmation concerning the impact of wildfires on air quality and PM2.5 concentration values, which is in accordance with the assumptions based on the research objectives.

This thesis employs the development of machine learning models in conjunction with the utilization of remote sensing data to efficiently evaluate PM2.5 concentrations, and the results are spatial distribution maps that can be reliable. In future research, it is recommended to undertake further refinement of the RF model's training process, which may entail the exploration of additional features or the fine-tuning of hyperparameters to enhance its predictive capabilities. Integration of additional data is another critical aspect for enhancing the robustness of PM2.5 estimation models.

-