結合土地利用迴歸與極限梯度提升演算法發展高雄都會區二氧化氮之推估模型

Development of an Integrated Model for NO_2 Variation Prediction in Kaohsiung Metropolis Using Land-Use Regression and XGBoost

翁佩詒(Pei-Yi Wong) ； 吳治達(Chih-Da Wu) ； 蘇慧貞(Huey-Jen Su)

《航測及遙測學刊》 26卷1期 (2021/03) Pp. 1-12

https://doi.org/10.6574/JPRS.202103_26(1).0001

摘要

暴露二氧化氮（NO_2）會對人體造成不良健康效應，然而過去空氣污染暴露評估模式仍有高估或低估的問題，因此使用高時空解析度之預測模型探討大範圍暴露濃度有其必要性。本研究以高雄都會區為研究區，使用土地利用迴歸模型、並結合極限梯度提升（Extreme Gradient Boosting, XGBoost）演算法，發展高時空解析度之NO_2濃度推估模型。結果顯示，土地利用迴歸結合XGBoost模型R^2為0.82，均方根誤差為4.53 ppb，具有高度預測與解釋力，十折交叉驗證R^2為0.82，顯示模型沒有過度擬合的問題，最後利用此模型推估高雄都會區NO_2之時空變異情形，結果發現高值熱點出現在南高雄之工商業發達以及人口密集處。

關鍵字

空氣污染 ； 二氧化氮 ； 土地利用迴歸 ； 極限梯度提升演算法

並列摘要

Exposing to Nitrogen Dioxide (NO_2) may cause adverse health effects. Previous air pollution estimating models still face overfitting or underfitting problems. Thus, using estimation model with high spatial and temporal resolution to assess NO_2 exposure is important. This study utilized Land-Use Regression (LUR) coupled with Extreme Gradient Boosting (XGBoost) algorithm to feature NO_2 concentration distribution in Kaohsiung metropolis. The results showed that R^2 value for LUR integrated XGBoost model was 0.82, RMSE was 4.53 ppb, which had highly explanatory ability. Besides, 10-fold cross validation R^2 for the proposed model was 0.82, which showed that the model did not encounter overfitting issue. Finally, this study used the model to depict estimation maps for NO_2 concentration variation in Kaohsiung. The results showed that higher polluted regions were clustered. in south Kaohsiung where industries were well developed and population was densely distributed.

並列關鍵字

Air Pollution ； Nitrogen Dioxide (NO_2) ； Land-Use Regression (LUR) ； Extreme Gradient Boosting (XGBoost)