臺灣博碩士論文加值系統

針對污染土壤中重金屬的濃度分佈具有變異大、高度偏歪和存在極端值的特性，本研究以地理統計法(geostatistics)並配合傳統統計技術來預測污染土壤中重金屬濃度的空間分佈，並研究如何達到最佳效益的採樣。
由於資料的變異性大，本研究比較不同樣本數以一般克利金法
(ordinary kriging)所得的預測值，結果顯示不同樣本數所預測的重金屬濃度空間分佈會有差異，原因在於極端值是否參與預測，欲以克利金法準確預測出重金屬濃度的空間分佈，則採取的樣本中必須包含極端值。半變異圖能表現資料的空間相依性，是克利金法的核心，本研究以不同的半變異圖模式來模式化實驗半變異圖，發現在高濃度的位置套用高斯模式所得的預測值會明顯比球面和指數模式所得的預測值高，尤其是參與預測的樣本數較少且包含極端值的情況。因為資料的偏歪和極端值的存在，與地理統計的假設不符並可能遮蔽了資料的空間結構，一般以自然對數轉換，但本研究結果發現自然對數轉換會使預測值產生偏差，為解決這樣的問題本研究以指標函數轉換，引用具有分佈自由(distribution free)特性的指標克利金法(indicator kriging)進行預測，結果指標克利金法所得的預測值比以自然對數轉換所得的預測值有較高的準確度。因為污染來源相同，使得不同種類重金屬濃度及表、土的重金屬濃度呈現良好相關，本研究應用Q-mode因子分析的步驟經適當的近似轉換，發現鉛濃度可有效地輔助鎘、鋅濃度空間分佈的預測，表土重金屬濃度可顯著地提升土重金屬濃度空間分佈預測的準確度。本研究為發揮地理統計在採樣決策的用途，以交叉驗證法(cross-validation)將資料區分為克利金法能準確預測與不能準確預測的兩部分，不能準確預測者即視為極端值，由前面試驗得知極端值是影響預測分佈準確與否的主因，故以是否取得極端值做為獲得足夠代表性樣本的指標，並考慮成本的因素，由極端值全部取得的機率與支出成本百分比隨樣本數變化的關係，可尋求滿足最佳採樣效益的最少樣本數，建立最有效的採樣決策。

The objectives of this study are using geostatistics and
traditional statistical methods to predict the spatial dis-
tribution of heavy metal in contaminated soils and to search
for the most efficient sampling strategy for contaminated site
characterization.Two study sites located at Tou-Yuan County
were selected.
The spatial distribution of soil Cd, Pb and Zn
concentrations were predicted by kriging in this study. The
results showed that the spatial distribution of those heavy
metal concentrations estimated by ordinary kriging varied with
the sampling numbers used. However, the accuracy of prediction
is mainly dependent on whether the outlier data are sampled.
There is no significant difference between the estimated
values using spherical and exponential semivariogram model.
But the estimated values using gaussian model were higher than
those using spherical and exponential models at the location
having high concentrations of heavy metal. The natural logtransformation of data of this study resulted in largees timated biasedness. The topsoil Cd and Pb concentrations
insite A predicted by indicator kriging were more accurate than
those predicted by natural log-transformed ordinary kriging.
The accuracy of topsoil Cd and Pb concentrations'' prediction can be improved by the assistance of other available Cd, Pb,and Zn concentrations of topsoils or subsoils using Q-mode factor approximation. The relationship between the probability of obtaining outliers and relative sampling cost is developed and proposed as a guideline for searching the most efficient sampling strategy for characterization of spatial distribution of pollutants in contaminated soils.