誤差分析平台運用於植生影像分類準確性提升之探討 －以台灣大學校區為例

Applying the error analysis platform in improving the accuracy of vegetation image classification -Take the campus of Taiwan University as an example

劉威廷(Wei-Ting Liou)

指導教授 ： 朱子豪

國立臺灣大學/理學院/地理環境資源學研究所/碩士(2008年)

https://doi.org/10.6342/NTU.2008.10243

摘要

現行土地利用調查之影像分類，有時需要超過95%之分類準確性，然而一般自動化影像分類後之準確性平均而言為85%左右，用人工判釋可達100%的準確性，但卻是極為耗費人力且不切實際的做法，因此如何在自動化分類後提升至要求的準確性即是本研究的主要目的。
　　本研究的誤差分析是利用像元經過最大相似法分類後，產生之類別機率值作為誤差判釋的基礎，像元類別機率值過低或過近是發生誤差的主要原因。因此研究方法運用誤差分析平台將像元的前三高類別機率值寫出，並經由閾值的設定找出可能發生誤差的像元，透過均質區的套疊，進行人機互動判釋以補足其準確性。
　　研究結果顯示在誤差分析概念上，像元類別機率值過低或過近的確是誤差發生的主要原因，但要注意影像的陰影與植生變異是降低分類準確性另一主因，且是不適用於誤差分析概念；在誤差分析平台上，透過閾值的設定，將有誤差類型的像元套疊至均質區中，經由人機互動方式補足準確性，是一種改進全自動與全人工之間－「全有」、「全無」概念的新做法。

關鍵字

分類準確性 ； 最大相似法 ； 誤差分析 ； 像元類別機率值

並列摘要

While the accuracy for interpretation of satellite images in land-use surveys is required to exceed 95% in some cases, automatic classification of such images can only reach an average accuracy level of about 85%. Although artificial interpretation can secure perfect accuracy of 100%, it is too labor-consuming to be practical. This thesis focuses on how to increase the accuracy for using auto-classification to a required level.
　　Using a maximum likelihood method, satellite images are automatically classified into several types of land cover, each with a computer-determined probability. For the probability for a type of land cover, the lower or the closer to the probability for another type of land cover, the more likely an error in automatic interpretation. This is an error analysis platform used in the thesis.
　　For an image, three types of land cover with the top three probabilities are selected for error analysis. For a type of land cover, if its probability is lower than a specified critical value or the difference between the probability and another probability (for another type of land cover) is lower than another specified critical value (the two probabilities are too close), artificial interpretation through overlapping picture elements on homogeneous areas is made to find whether the auto-classification is correct.
　　Using the error analysis, it is found that too low probabilities for types of land cover or too close probabilities between two types of land cover are the main reason behind the interpretation error due to auto-classification. However, shades in images or variation in vegetation also account for such errors and the two factors are not suitable for being incorporated into the error analysis.

並列關鍵字

accuracy ； Maximum Likelihood classifier ； error analysis ； pixel probability

參考文獻

林孟龍 (2004) 以景觀層級界定生物多樣性保育範圍－中尺度資源衛星MODIS影像的應用，臺灣大學地理環境資源學研究所博士論文。

行政院農委會農糧署 (2007) 應用高時間與空間解像力遙測影像於水

邵泰璋 (1998) 類神經網路於多光譜影像分類之應用，交大土木工程系碩論。

鄧東波、丁亞中、林裕彬 (2001) 利用高解析度衛星影像於都市綠地空間分佈之研究，中華地理資訊學會2001年年會暨學術研討會。

Congalton, R. (1991) A Review of Assessing the Accuracy of