高光譜主軸轉換影像辨識土地利用型最適主軸數決定方法之研究

Study on Deciding the Optimal Number of PCA Components of Hyperspectral Images for Land Use Classification

林金樹(Chin-Su Lin)

《台灣林業科學》 17卷4期 (2002/12) Pp. 471-481

https://doi.org/10.7075/TJFS.200212.0471

摘要

本研究利用都會區及濱海區高光譜影像已知的土地利用型資料，探討應用主軸轉換不同主軸數組合影像辨識土地利用型資訊之效度問題。研究結果顯示，在監督式最大概似法分類法則與相同土地利用型組類及分類凖確度指標等共同的評估基礎下，特性根閾值法、特性根陡坡圖檢定法以及累積變異數百分比閾值法等傳統的決定主軸數方法，並無法同時滿足降低資料維度與保持分類準確度的目標，而本文所提出的相鄰特性根斜率閾值法比較能掌握土地利用型分類之效度，應用該法所決定的主軸數組合影像，對都會區及濱海區兩組試驗影像之土地利用型分類準確度的損失量均低於2%，同時可以節省兩倍及二十倍的土利用型分類時間。應用主軸轉換影像檢測土地利用型時，必須注意增加主軸數可以提昇土地利用型細部組類的辨識能力。

關鍵字

遙測 ； 高光譜影像 ； 主軸轉換法 ； 土地利用型 ； 相鄰特性根斜率閾值法

並列摘要

This study used 2 hyperspectral data sets that cover well-known land use areas, 1 urban and 1 marine site, to explore the validity of land use recognition by varying different numbers of principal components of the images. Validity is defined as the overall accuracy of the classification under the same classification algorithm and same set of land uses for each site on the experimental image. Result shows that the traditional methods of the threshold of eigenvalue (ET), the scree test (ST), and the threshold of cumulative variance proportion (TCVP) chose PC numbers which could not satisfy the data dimensionality reduction while main taining a relatively higher validity of land use recognition. In contrast, the threshold of the slope of adjacent eigenvalues (TSAE) method announced in this paper could be used to effectively determine the suitable principal components for land uses recognition. The overall loss of accuracy for those detailed land uses in hyperspectral images of both urban and marine sites were less than 2%, and the classification operation was reduced 2-and 20-fold of the cases of the classification with all original bands. Additional PC components can improve the ability to recognize detailed land uses. Researchers should keep it in mind when using the PCA method for land use classification.

並列關鍵字

remote sensing ； hyperspectral date ； principal component analysis ； land use ； TSAE method