逐層分割群聚法及反覆移動均值群聚法於地覆非監督式分類之比較

Comparison between Hierarchical Partitioning and Iterative Migrating Means Clustering Algorithms in Unsupervised Classification of Land Cover

黃凱易(Kai-Yi Huang)

《航測及遙測學刊》 5卷3期 (2000/09) Pp. 43-55

https://doi.org/10.6574/JPRS.2000.5(3).3

摘要

群聚在法常用於執行地覆非監督式分類、顯示資料內部結構及協助監督式分類選取所需之訓練樣本。常用的反覆移動均值法，有數項輸入參數，然群聚結果易受輸入值左右，且因需遞迴運算，效率較差。本研究之目的係發展一逐層分割群眾法，可改善上述問題，作為反覆法的替代性選擇。逐層分割法是依據超平面和二元樹建構而成。本研究以實驗田園錄影資料檢測兩群聚法，並就操作難易、運作效率、參數影響及分類準確度等條件比較結果。逐層分割法只需輸入一個參數，即最大群集數，且容易操作，受輸入參數之影響較小。逐層分割法效率遠高於反覆移動均值法，惟在分類準確度上稍遜於反覆法，然可能純屬統計誤差。因此逐層分割法在地覆非監督分類上確可為反覆移動均值法的替代選擇。

關鍵字

群聚法 ； 逐層分割群聚法 ； 反覆移動均值法 ； 非監督式分類

並列摘要

Clustering algorithms are commonly used to perform an unsupervised classification, to expose internal data structures, and to facilitate the selection of training samples for a supervised classification. The iterative migrating means clustering algorithm (IMMCA) has several input parameters, some of them are difficult to determine, and thus clustering results are easily affected by inappropriate values of input parameters. Moreover, IMMCA is inefficient due to its iterative clustering process. The objective of this study was to develop a hierarchical partitioning clustering algorithm which can alleviate the problems stated above and serve as an alternative to IMMCA. The clustering algorithm was developed based on the concepts of hyperplane and binary tree. The two clustering algorithms were compared in the unsupervised classification of the video image data in terms of friendliness, efficiency, influence of input parameters, and classification accuracy. The new algorithm needed to determine only an input parameter, the maximum number of cluster, and was users friendly. It was much more efficient than IMMCA, but had a slightly lower classification accuracy than that of IMMCA. Therefore, it could be used as an alternative to IMMCA when applied to an unsupervised classification of land cover.

並列關鍵字

Clustering algorithm ； Hierarchical partitioning algorithm ； Iterative migrating means algorithm ； Unsupervised classification