應用主軸轉換法辨認空載高光譜影像土地利用型特性之研究

Application of the PCA Method to Land Use Characterization Using Airborne DAIS Hyperspectral Data

林金樹(Chinsu Lin)

《台灣林業科學》 17卷3期 (2002/09) Pp. 347-359

https://doi.org/10.7075/TJFS.200209.0347

摘要

本研究以台中市都會區之空載DAIS-3715高光譜影像為材料，探討應用主軸轉換法(PCA)檢測土地利用型的效能。原始的空載高光譜影像經過二階多項式橡皮伸張法修正後，所得修正後影像的像元大小為6m，其地面控制點與檢核點的平均誤差值小於3m，總體的幾何誤差小於半個像元。研究結果顯示，DAIS-3715原始24個波段的輻射亮度影像之光譜資訊，有99.87%被集中於PCI至PC5等五個主軸，以水體、道路、建築物、人工鋪面、行道樹、草地等六種土地覆蓋型的55個土地利用型組類為檢測目標時，五個主軸影像的分類準確率可達80%以上，比訊號雜訊比SNR指標及影像參數加權統計法PN指標所選出的5個具有較佳影像品質的原始波段影像的分類準確率高出20%；若以土地覆蓋型為檢測目標時，則PCA法可達98%以上之準確率，比SNR及PN法高出約10%以上。應用PCA法於高輻射解析力遙測資料之土地利用型檢測時，必須注意維持使各個主軸影像具有與原始影像相同的輻射解析力，方可發揮PCA有效降低高光譜資料維度的功能，並避免發生主軸資訊減量的現象。

關鍵字

空載高光譜影像 ； 土地利用型 ； 土地覆蓋型 ； 主軸轉換法 ； 遙測

並列摘要

This paper examines the efficiency of integrating PCA and MLC methods for detailed land use characterization using DAIS-3715 hyperspectral images. An urban area located in Taichung City, west-central Taiwan was selected for this research. Airborne DAIS hyperspectral images were geometrically corrected using a 2(superscript nd) -order polynomial model. The spatially residual mean squared error (RMSE) of the control GCPs and check GCPs for the rubber sheet stretching method were calculated as being less then 0.5 of a pixel, i.e., 3m. Results showed that totally about 99.87% of the spectral information of all 24 bands of the geo-corrected DAIS-3715 radiance image was transformed into 5 major components of PCl to PC5. Use of this 5-PC image could achieve an overall accuracy of 80% in recognizing detailed land use, for 55-classes of 6 major land covers, i.e., water, roads, buildings, man-made surfaces, vegetation, and grass. Classification of the same land use using the 5-band images with the best signals as determined by the SNR and PN method achieved only a 60% classification accuracy. Also at the level of land cover recognition, the 5- PC images could identify the 6 types of land cover very well with an accuracy of 98%. which was 10% better than that achieved by SNR and PN.
Results indicate that the PCA method can effectively gather spectral information of all bands of original radiance images into several components to achieve a much better accuracy for both land use and land cover detection. It was also proven that an effect of the component information reduction (CIR) effect might occur in the PCA images if a higher radiometric format (16 bit) of the original hyperspectal image was transformed as a lower radiometric format (8 bit) PC image. In considering both the efficiency of data dimensionality reduction and land use recognition, researchers should be careful to avoid the CIR effect in hyperspectral image processing.

並列關鍵字

airborne DAIS-3715 hyperspectral data ； land use ； land cover ； PCA ； remote sensing