以自組織映射圖及粗糙集理論判釋萬大水庫崩塌地

The Study of Image Classification on a Landslide Area through Self-Organization Map and Discrete Rough Sets

萬絢(Shiuan Wan) ； 閻嘉義(Jia-Yih Yen) ； 劉怡麟(Yi-Lin Liu)

《中華水土保持學報》 42卷3期 (2011/09) Pp. 240-250

https://doi.org/10.29417/JCSWC.201109_42(3).0007

摘要

過去多以統計方法分析空間資訊，以地形、植生、地貌等影像資訊為基礎，判釋各類崩塌地。現在的防災科技，更容易取得大量的屬性資料，然而資料量的增加不代表判釋能力的增加，且於資料量大、重覆且不確定的資訊而造成誤判。影響坡地崩塌主因在植生覆蓋以及暴雨地震等相關因素。然而暴雨地震屬事件型因子難以掌控，有鑑於此，本研究使用衛星影像原始光譜與DEM的高程等相關資料，並計算坡度高程與植生指標作為影像整合分類資訊，再用(a)自組織映射圖(self-organizing map, SOM)及(b)粗糙集理論(discrete rough set, DRS)，作為影像空間特徵分類方法，以達快速判釋崩塌地，有效降低時間及人力之目的。實驗區在萬大水庫，選擇了30筆崩塌地樣本的光譜值，第一步先就高程和坡度等資訊，建立門檻值，再以上述(a)、(b)兩種方法，以光譜值與八個植生因子分別進一步判釋崩塌地與非崩塌地。我們設計了一個以坡地門檻值配合光譜資訊的改善崩塌地判釋方法，結果顯示，以兩種影像空間特徵分類方法，在使用未加入坡地門檻值資訊時，其正確率皆可達到78%，但經考量門檻值後，則增進了大約10%。

關鍵字

自組織映射圖 ； 粗糙集理論 ； 植生指數 ； 坡地門檻值

並列摘要

Spatial information of a landslide is usually analyzed by statistical analysis based on the topography, vegetation, and landform information. However, due to the uncertainty and the massive amount of data, the determination will result in many misjudgments.This paper presents a multi-category image classification. The slope of the DEM, normalized difference vegetation index, and other modification of spectrum information are used. The (a) self-organizing map and (b) Discrete Rough Set, are then used to achieve rapidly interpret the landslides. We selected 30 samples of landslide spectral values surrounding the Wan-da reservoir. Following (a) and (b), the classifier was used to distinguish landslide and non-landslide areas. The results showed the spatial characteristics of the two image classification methods. We also examined eight different indicators of vegetation.We designed a novel method to effectively improve performance classification using the threshold method. Regardless of the slope thresholds, the accuracy could only reach 78%. However, if the thresholds are considered, the accuracy can be improved by about 10%.

並列關鍵字

self-organizing map ； rough set theory ； vegetation index ； slope threshold