自動化時空過程推估方法之發展及應用，並以台北都會區空氣懸浮粒子時空分佈之研究為例

Abstract

許多地理統計方法都假設資料隨機過程是定常性(stationary)與同質性(homogeneous)。然而在環境資料隨機過程中(例如:PM的時空分佈)，常常不具定常性及同質性，也尌是存在著趨勢(trend)。本研究發展一個自動化讓非定常性(non-stationary)及非同質性(non-homogeneous)隨機過程資料的時空共變異數轉換成定常性及同質性資料的時空共變異數。本研究分別使用Kernel smoothing及粒子群最佳化演算法(particle swarm optimization method, PSO)和Nelder-Mead單體法(Nelder-Mead simplex method)來計算趨勢及參數的估計。並利用這些方法來迭代以求得最佳的趨勢與擬合時空共變異數(covariance fitting)。 許多研究指出細懸浮粒子比粗懸浮粒子更容易進入人體造成危隩並影響生態。估算對人類及生態的衝擊需要長期的暴露資料，但是在台北都會區過去並無系統的監測細懸浮粒子，直到2005年8月整個監測網路才完成。台北都會區主要的污染源以工業及交通為主。懸浮粒子相關資料(如:PM10、PM2.5和TSP)獨立搜集於中央及地方政府。在本研究使用貝氏最大熵法(Bayesian Maximum Entropy method, BME)去整合(a)時空的懸浮粒子(b)特殊位置上懸浮粒子的確定性資料(hard data)與不確定性資料(soft data)(c) PM2.5/PM10 and PM10/TSP 比值關係去回推過去2003-2004年台北都會區PM2.5時空分佈機率密度函數(Probability Density Function, PDF)並與觀測值做比較。 本研究利用所提之自動化方法來推估台北市PM10時空分佈中之趨勢與共變異數模式之最佳化參數。PM2.5之回推預測結果顯示本研究可提供合理之推估結果，以2003年新莊超級測站及2004年PM2.5測站為例，其推估相對誤差於分別為10.6%與10.7%，分析結果顯示較高的PM2.5、PM2.5/PM10及PM2.5/TSP值發生在大同區、中山區偏南、中正區及新莊地區。

Mary geostatistics approached assume the homogeneity and stationarity of the data process. However, the assumption is not valid for most if environment processes of interest (e.g. spatiotemporal distribution of PM). This study developed an automatic scheme to discompose a nonstationary and nonhomogeneous process into a deterministic trend and a random process which can be characterized by the stationary and homogeneous S/T covariance model. Kernel smoothing method and particle swarm optimization method as well as Nelder-Mead simplex method were applied for trend modeling of parameter estimation, respectively. By the proposed scheme, the spatiotemporal bandwidths as well as the covariance parameters are estimated iteratively in order to account for the goodness-of-fit of trend and covariance modeling as well as the complexity of nested covariance model and S/T correlation among the dataset. Numerous studies have shown that fine airborne particulate matter particles (PM2.5) are more dangerous to human health than coarse particles, e.g. PM10. The assessment of the impacts to human health or ecological effects by long-term PM2.5 exposure is often limited by lack of PM2.5 measurements. In Taipei, PM2.5 was not systematically observed until August, 2005. Taipei is the largest metropolitan area in Taiwan, where a variety of industrial and traffic emissions are continuously generated and distributed across space and time. PM-related data, i.e., PM10 and Total Suspended Particles (TSP) are independently systematically collected by different central and local government institutes. In this study, the retrospective prediction of spatiotemporal distribution of monthly PM2.5 over Taipei will be performed by using Bayesian Maximum Entropy method (BME) to integrate (a) the spatiotemporal dependence among PM measurements (i.e. PM10, TSP, and PM2.5), (b) the site-specific information of PM measurements which can be certain or uncertain information, and (c) empirical evidence about the PM2.5/PM10 and PM10/TSP ratios. The performance assessment of the retrospective prediction for the spatiotemporal distribution of PM2.5 was performed over space and time during 2003-2004 by comparing the posterior pdf of PM2.5 with the observations. By the proposed scheme, the optimal parameters of trend and covariance models are obtained from PM10 dataset. The retrospective predictions of PM2.5 provide reasonable results in which the relative errors in 2003 at Sinjhuang and 2004 at TWEPA stations are 10.6% and 10.3%, respectively. High values of PM2.5 concentration and the ratios of PM2.5/PM10 and PM2.5/TSP are shown in the areas of Datong district, south of Jungshan district, Jungieng district and Sinjhuang.