寒帶沼地高光譜影像分類之研究

Abstract

寒帶沼地高光譜影像分類之研究 學生：郭麟霂 指導教授：史天元 國立交通大學土木工程學系 摘 要 寒帶沼地之面積雖只佔全球面積不到百分之四，但由於沼地之氧化作用，產生了大量的二氧化碳。因此該區域之環境變遷對全球之氣候變化影響甚鉅。其中該區之地面覆蓋分類圖為其他環境變遷研究之重要參考數據之一，本研究針對寒帶沼地使用高光譜影像進行分類研究。 如何增加類別分離度、降低資料維度及選擇一個較適合的分類方式為分類成果良窳之重要因素，此外如何將遙測影像高維度資料有效地經由轉換，達到降低資料維度、提高處理效率及去除波段之間的高相關，亦相當之重要。本研究採用低通濾波(Low Pass Filter)及主軸轉換(Principle Component Analysis, PCA)和MNF(Minimum Noise Fraction)轉換欲獲得最佳之頻譜組合，並使用最短距離，最大似然及光譜角映射(Spectral Angle Mapping, SAM)等分類法來分析比較其對分類結果之影響。 根據實驗的成果發現，以原始影像進行低通濾波可有效提升分類精度，但必須以波段數足夠為前提。因此若能配合MNF轉換或主軸轉換，則可使用較少之波段數即可得到較高之分類精度成果，就最後分類精度而言，MNF轉換優於主軸轉換。而分類法方面，高斯最大似然分類法明顯優於最短距離分類法及光譜角映射分類法。

A Study on Hyperspectral Image Classification of Boreal Area Student: Lin-Mu Guo Advisor: Dr. Tian-Yuan Shih Institute of Civil Engineering National Chiao Tung University Abstract Boreal forests cover less then 4% of the total area of the earth, but the oxidation inside marshes produce lots of carbon dioxides, this might cause a global climate change, such as the greenhouse effect. As a result, the ground coverage map of boreal area becomes an important reference data for other researches. This study applies a frame of AVIRIS image of boreal area for land cover classification. To increase the separability, and reduce the data dimensionality, choosing a good classifier and removing the high correlation between bands are important factors that affect the result of hyperspectral image classification. In this study, LPF (Low Pass Filter), PCA (Principle Component Analysis) and MNF (Minimum Noise Fraction) are applied to obtain the best spectral combination for classification. The results from different classification approaches are then compared. According to the result of this research, Low Pass Filter processing can improve the classification accuracy effectively. But combining with MNF or PCA transformation can obtain higher accuracy in lower dimensionality. In general, MNF transformation is better then PCA transformation. Regarding to classification method, Maximum Likelihood Classifier performs much better then Minimum Distance Classifier or Spectral Angle Mapping classifier.