近年來，空氣污染為許多國家嚴重問題，其中人們最關心的是PM2.5對健康的影響。目前，台灣約有80個監測站提供PM2.5的濃度監測，且主要集中在西部地區，然而，中心點監測儀的測量通常缺乏足夠的空間和時間分辨率以描述研究人群的暴露變化。因此，本論文主要致力於加以考量氣象觀測資料、除PM2.5之外的空氣污染物濃度、交通與土地利用等可能影響PM2.5濃度之因素，與相關領域專家之討論後，進行合理且適切的資料篩選與處理，接著，利用線性回歸、決策樹與隨機森林等機器學習方法建立估計台灣無測站區域的即時PM2.5濃度模型，其中以隨機森林方法建立之估計模型最為精確，其十次交叉驗證得到的評估指標決定係數（R2）的值為0.651、均方根誤差（RMSE）與平均絕對誤差（MAE）分別為11.05μg/m3與7.87μg/m3，相較先前採用神經網路或迴歸模型之研究，誤差值相去不遠。最後，我們將此模型應用在即時的PM2.5的濃度推估並呈現在網頁上，提供地區PM2.5濃度的查詢與每小時全台灣PM2.5濃度的分佈情形。;Recently, air pollution has become a serious problem in developing countries. Among them, people are most concerned about the impact of PM2.5 on health. In Taiwan, about 80 stations provide PM2.5 concentrations monitoring, which is mainly concentrated in the west. The location of the monitoring station is mainly set up at various agencies or school sites and roofs. In addition, the measurements from a central point monitor often lack sufficient spatial and temporal resolution to capture the exposure variability of the study. In this study, we collected the meteorology, air pollution, traffic-related and land use data. After data screening and processing, we estimated hourly PM2.5 concentrations in ungauged sites in Taiwan by using linear regression, decision tree and random forest. Random forest has the best performance in estimating PM2.5 concentrations. We achieve the value of 10 cross-validation coefficient of determination (R2), root mean squared error (RMSE) and mean absolute error (MAE) are 0.651, 11.05μg/m3 and 7.84μg/m3 respectively. Compared with previous studies using neural networks or regulatory models, the error between our estimation and the actual measured value is not large. We also applied the model to the real time PM2.5 concentrations estimation and showed it on the website, which can provide the query of PM2.5 concentrations in the region and the distribution of PM2.5 concentrations in Taiwan every hour.
