衛星遙測影像之環境資訊萃取

Environmental Information Extraction from Satellite Remote Sensing Imagery

江介倫(Jie-Lun Chiang)

指導教授 ： 鄭克聲

國立臺灣大學/生物資源暨農學院/生物環境系統工程學研究所/博士(2005年)

https://doi.org/10.6342/NTU.2005.00099

摘要

環境資料可概分為時間資料、空間資料、時空資料等三類。例如一氣象站之溫度記錄屬時間資料；獲取單幅衛星影像屬空間資料；而若為多時段，甚或一序列的影像資料則既涵蓋了空間資料也包括時間資料。由於衛星遙測資料具有廣景攬要、即時監控的優點，且近年來衛星影像品質與解析度大幅提昇，故廣為運用於自然資源、環境及災害等之監測。如何由龐大的資料記錄有效萃取出所需資訊是極為重要的課題。本研究主要探討由衛星影像萃取所需資訊的方法，其中涵蓋：1.影像分類 2.影像融合3.變遷偵測。文中將依序探討這三大主題。
影像分類主要是在特徵空間中劃分出各類別所佔據的空間；以常用的高斯最大概似法為例，它須假設各類樣本點在特徵空間中呈常態分佈，雖資料非屬常態分佈時，仍常被採用，但其分類精度會降低。故我們提出一無母數的方法-指標克利金法，它不需假設各類特徵之機率分佈型式；此法將各種類別在特徵空間之分佈視為一個指標變數之隨機變域，並發展一個以指標變域空間推估為基礎之新分類理論；它推估特徵空間中每一點屬於每一類別的機率，最後再將此點歸為具有最大機率的類別。此分類法可處理特徵空間中非等向分佈之問題，且無兩種以上類別之樣本點對應到特徵空間中同一點時，其對訓練像元之分類正確率可達100%。
本研究中，以台北地區的SPOT衛星影像為例，比較本研究所提出的指標克利金分類法與傳統常用的最大概似法之分類結果；結果顯示即使是針對驗證樣本，指標克利金分類法亦可達到99.6%之分類正確率，無論在訓練樣本或驗證樣本，皆高於最大概似法之分類正確率；將分類結果與該區相近時期之航空照片比較，亦顯示指標克利金分類方法在整個區域之分類表現明顯優於最大概似分類法。
影像融合主要在結合不同來源影像，截長補短，提高影像解析度，甚至從中獲取無法單獨由任一影像得到的資訊。本研究中首先以預分類獲取各類地表覆蓋光譜特性知識，再配合尺度轉換，建構一新融合技術，稱為知識基礎之尺度轉換融合，在此方法中空間資訊可在尺度轉換的過程中被融合進新影像，而光譜資訊可以由預分類的過程中，保留於各類別中，藉此我們可以得到兼具多光譜資訊及高空間解析之新影像。
本研究中並以SPOT衛星之多光譜與全譜態影像融合，得到較高空間解析度之多光譜影像，比較本論文中所提出之融合方法所得的融合影像與影像線性疊加、影像相乘等其他空間域融合方法之融合影像，結果顯示知識基礎之尺度轉換融合技術確實能得到較高解析度之多光譜影像，且對於如街廓或學校操場等地徵的呈現較其他方法更清晰、明確。由此可知本研究提供空間域影像融合一項新的選擇，此方法可有效的結合高空間解析的資訊與高光譜解析的資訊，得到一多光譜的高解析度影像，使假色影像清楚呈現更多細節，以利影像後續之判識或應用。
變遷偵測最常用的方法為比較不同時期影像之差異，並訂定一門檻值以決定多大的差異程度以上，將之視為變遷。而一般研究中，門檻值的訂定較為主觀；且評估結果時，地表變遷前後皆須有變遷位址之地真資料方可如評估分類精度般，以混淆表評估之，受限於此，評估變遷偵測結果實屬不易；因此本研究假設前後兩時期影像構成之散佈圖為雙變數常態分布，並基於假設檢定的概念，以某點非為變遷為虛無假設，拒絕虛無假設之像元視為變遷；亦即分析各類別在兩時期之灰階值皆屬同類之聯合機率，並劃定不同信賴區域以外之離群值為變遷。如此在偵測出變遷點的同時，可以提供其為變遷之信賴水準，以代替混淆表中生產者正確率之功能。
本研究回顧衛星影像之分類、融合與變遷偵測等應用主題，並分別提出新的方法，其中地理統計、尺度轉換、預分類、假設檢定等概念與理論被用於其中，使由衛星影像萃取所需時空資訊時能更有效與精確；透過這些分析方法，我們可以有效的由大量的衛星遙測資料中去蕪存菁，萃取出我們所需要的環境資訊，增加衛星遙測資料在環境、資源監測的應用潛力與效率，且能使衛星遙測技術可以更廣泛的應用於環境課題。

關鍵字

影像分類 ； 指標克利金 ； 影像融合 ； 尺度轉換 ； 預分類 ； 知識基礎 ； 變遷偵測 ； 假設檢定 ； 信賴區域

並列摘要

Environmental data can be classified into three categories: spatial data, temporal data, and spatio-temporal data. For example, a remote sensing image can be considered as spatial data; daily temperature observations of a weather station are temporal data, and multi-temporal images or a sequence of remote sensing images can be considered as spatio-temporal data. In recent years, the number of remote sensing satellites with high pixel resolution has increased significantly and remote sensing images were widely used for environmental monitoring and related applications. Therefore, extraction of useful information from satellite images acquired by various sensors is crucial for the success of such applications.
The main objective of this study is to develop techniques for extraction of land-surface information from satellite remote sensing images. Implementation of such techniques is demonstrated by three applications - image classification, image fusion, and landcover change detection.
Image classification
A nonparametric indicator kriging (IK) approach of remote sensing image classification is developed. The work of image classification is transformed into estimation of class-dependent probabilities in feature space using indicator kriging. Each pixel is then assigned to the class with maximum class probability. The IK classifier yields 100% classification accuracy for training data provided that colocated data in feature space do not exist. An example of landcover classification for the Taipei metropolitan area using SPOT images demonstrated that the proposed indicator kriging classifier is superior to the maximum likelihood classifier in terms of producer’s and user’s classification accuracies.
Image Fusion
A knowledge-based scale transfer (KBST) fusion technique was developed. Firstly, SPOT multispectral (XS) images were used for major categories (water, vegetation, and bare soil/built-up) landcover classification of the study area. Regression relationships between digital numbers of panchromatic (PAN) and XS images were then established and used for subsequent scale transfer. The class-specific regression models help to preserve spectral information during scale transfer. Finally, a scale transfer algorithm using class-specific regression models was adopted for fusion of SPOT XS and PAN images, yielding a multispectral, high-resolution image which offers more details of the study area than other spatial domain fusion techniques.
Change detection
A hypothesis-test approach for landcover change detection was developed using multispectral images acquired in two different dates. Major category landcover classification using multispectral images was implemented for both (pre- and post-) images to identify sets of no-change pixels. Using only no-change pixels, same-band digital numbers of the two images of specific landcover classes were assumed to form a bivariate normal distribution.
The work of change detection was then placed in the framework of hypothesis test with null hypothesis of no-change. Critical regions with respect to a chosen level of significance a were then determined for each landcover class using the bivariate normal distribution. Finally, landcover changed areas were determined with desired confidence levels. We demonstrated that the hypothesis-test approach of change detection is promising and deserve further investigation.

並列關鍵字

Image Classification ； Indicator Kriging ； Image Fusion ； Scale Transfer ； Pre-classification ； Knowledge-Based ； Change Detection ； Hypothesis test ； Confidence Region

參考文獻

[1] Bates, P.D., Horritt, M.S., Smith, C.N., Mason, D. (1997), “ Integrating remote sensing observations of flood hydrology and hydraulic modelling”, Hydrological Processes ,11: (14) 1777-1795.

[2] Bendediktsson, J.A., Philip H. Swain and Okank Ersoy, (1990) “Neural Network Approaches Versus Statistical Methods in Classification Remote Sensing Data”, IEEE Transactions on Geoscience and Remote Sensing, 28(4), pp.540-551.

[3] Bischof, H., W. Schneider, and A. J. Pinz, (1992). “Multispectral Classification of Landsat-Images Using Neural Networks”, IEEE Transactions on Geoscience and Remote Sensing, Vol.30, No.3, pp.482-490.

[4] Brekke, Camilla, Anne H. S. Solberg, (2005). “Oil spill detection by satellite remote sensing”, Remote sensing of Environment, 95(1), pp. 1-13.

[5] Chen, K. S., Y. C. Tzend, C. F. Chen, and W. L. Kao,(1995) “Landcover Classification of Multispectral Imagery Using a Dynamic Learning Neural Network”, Potogrammetric Engineering and Remote Sensing, Vol.61, No.4, pp.403-408.