空間資料探勘與知識產生-以建立崩坍敏感性評估模式為例

Abstract

台灣的山地集水區在各項環境因子的綜合影響下，使得山崩發生的頻率與地點偏多，為了保障人們的安全，關於崩坍地的調查、評估與預測之研究上，產出甚多，尤其隨著GIS的應用，許多模式評估的成果皆有效且相當顯著；然而，隨著大量資料的累積與知識的不斷推陳出新，如何利用舊有的知識來產生快速新的知識，以提供我們對崩坍的機制甚至是時空分佈上特性上更進一步的瞭解，是迫不及待的事；因此，研究希望利用空間資料探勘技術來分析環境的崩坍特性，藉以更深入瞭解造成崩坍發生的基本環境因素為何，進一步產生崩坍及其相關因子之間的知識描述。 研究首先對過去相關的崩坍研究與空間資料探勘進行文獻回顧，形成研究方法，並挑選建立模式之因子，接著利用空間統計與決策樹分析的方法，建立並歸納環境因子與邊坡崩坍之間的關係模式，再以法則的方式來描述崩坍產生的環境特性，然後將模式模擬的結果，寫入GIS中，以對崩坍地資料進行分類，評估模擬結果，最後提出結論及建議。 在結果方面，利用決策樹產生出來的四個分類法則對研究區的資料進行崩坍地預測，發現成功度最高者可達88.46%，最低也有73.08%，可見利用這樣的方式作崩坍潛勢評估，效果不錯，是一個可行的辦法。 關於後續研究，因為資料的正確度與精度對崩坍模擬而言是非常重要的，因此若能加入更新的資料，或者結合不同的分類法則，應該可以更提高模式分類及預測精確度，以供未來研究或評估使用。

In Taiwan, because of the influence of each environmental factors, landslides occur frequently around the basins. To ensure the safety for people residents, there are more and more researches to investigate, estimate, and predict landslides via the application of GIS. However, along with the process of accumulation of data and knowledge, it is urgent to employ existing knowledge to generate new knowledge to provide with further understanding of the machanisms and spatiotempotral distribution of landslides. For this reason, this research utilizes the technology of Geospatial Data Mining(GDM) to obtain relationships between the landslides and the environmental factors to deeply understand the landslides. Firstly, the methodology is establshed by carrying out review of research of landslides and Geospatial Data Mining. Afterwards, the factors of model is selected, the method of spatial statistics and decision tree is built, the relationship between the environmental factors and landslides is summed up, and the environmental characters of landslides are described with the form of regulations.. This research develops a program of Arc Object to classify landslides data. Finally, the simulation results are evaluated and the conclusions and suggestions are proposed. The results show that the highest accuracy rate of the predicting model reaches to 88.46% while the lowest one is 73.08% by using the four classifications generated by decision tree to classify the data to predict the distribution of landslides in the area under investigation. In the future, since the validity and precision of data of assessing landslide susceptibility is very important, the result of simulation can be improved by updating the data or combining several classifications in the model.