| 摘要: | 應用影像進行自動化變異偵測時，因影像特性或應用需要的不同可分為單時性(uni-temporal)以及多時性(multi-temporal)影像變異偵測。在本研究中，將針對單時性以及多時性衛星影像之變異偵測分別提出對應的處理方法以及分析結果。
在單時性衛星影像變異偵測方面，實際上其概念是以空間變異(Spatial Change)為出發點進行分析。理論上在發生變異的範圍相對於未變異之區域為有限時，可將影像上有變異像元視為異常物並加以偵測。因此在本研究中採用異常物偵測(Anomaly Detection)方法為基礎，並配合EM (Expectation-Maximization)方法與貝氏理論(Bayes Theory)，在原始影像中將未變異之背景與發生變異之異常物加以分離，以達到變異分析的目的。此外在研究中由於使用EM方法分析非變異與變異像元之理論機率分布，因此在得到變異結果後，可同時推估其精確度理論值，並且進而應用所提空間過濾方法於成果之進一步改善。例如在本部份研究中以應用光學影像分析海洋油汙染做為測試，實際應用本研究所發展之方法來偵測海面上因油汙染而產生變異區塊時，針對分析所得汙染物之錯誤率(False Alarm)可達到10%以下。並且若用空間過濾方法進一步改善成果時，其錯誤率可再降低至5%以下。
而在多時性衛星影像變異偵測方面，主要是應用於多時期衛星影像間之時間變異(Temporal Change)分析上。而進行此類分析時，其必須進行的程序則主要分為多時性影像灰值正規化與影像變異偵測兩項工作。關於多時性影像灰值正規化上，其進行之原因為不同時期影像在獲取時會因其相關環境因素之改變，造成不同時期影像在輻射上有顯著之差異。通常造成此差異之主要因素是來自於大氣以及感測器之成像記錄過程有關，而影像輻射校正即為克服此一問題之處理方法。但絕對影像輻射校正通常需要極高之成本，對於大範圍之變異偵測在時效上與應用有其限制。同時由於多時性之影像變異偵測所需的影像灰值正規化理論上不需進行絕對性之校正，因此在大多數之情形僅進行相對性之校正即可。而關於多時性變異偵測上，則須經由比較兩時期影像之差異後再經由門檻值之設定後才能得到變異之結果，然而此門檻值通常不是定值，並且會受到影像品質的影響而有所改變，甚至會因影像輻射校正之模式不同而有所差異。在本研究中，關於相對影像輻射校正的部分主要應用虛擬未變異特徵(Pseudo Invariant Features, PIFs)的概念進行。在過程中先由原始多時期影像萃取PIFs之後，再利用PIFs進行多時期影像之相對校正。而在多時性影像變異偵測上，在本研究中將延伸PIFs萃取之概念，發展萃取虛擬變異特徵(Pseudo Variant Features, PVFs)的方法，在影像中進行PVFs之分析。在取得PVFs之後，即可參考所得之PVFs並配合變遷向量(Change Vector)之計算來推估影像中可能發生變異之區域。在實際應用上，本部分的研究採用不同感測器所獲取之多時期光學影像進行不同區域之土地變異之分析。在成果經過檢核後，不同區域之變異偵測在精確度上均可達到90%左右，同時其Kappa Coefficient亦可達到0.8左右。
實際上單時性以及多時性影像變異偵測之主要差異在於變異程度資訊之計算方式不同，例如在單時性影像是以影像的背景統計為基礎，進而推算影像上每個像元相對於背景之異常程度。而在多時性影像上，則是在輻射校正後針對不同時期影像之光譜差異來分析變異發生之程度。然而就變遷偵測而言，其中最重要的部分則是如何取得變異以及非變異參考資訊。因此本研究在單時性以及多時性影像的變異偵測中，均提出相對應的變異以及非變異參考資訊的推估方法，理論上除了將其應用於本研究所提出的變異偵測過程之中，未來亦可應用於其他的需要參考資訊的變異偵測方法上。
Theoretically, change detections using remotely sensed images can be categorized into uni-temporal and multi-temporal image change detections according to the image properties and practical applications. In this study, the methods for the two categories of change detections using images of optical satellite sensors were proposed and applied to test their performance.
The uni-temporal image change detection concept is based on the extraction of spatial changes on the source image. Assuming that the area of change is limited when compared to the whole study area, the changed pixels on images can be regarded as an anomaly and detected
therefore, this study applied the theory of anomaly detection to detect specific spatial changes. To model and separate unchanged background and changed anomaly, the Expectation-Maximization (EM) theory and Bayes theory were applied. In addition, EM can be used to estimate the probability distribution functions (PDFs) of unchanged background and changed anomaly. Using the estimated PDFs, the theoretical accuracy of detected changes can be further estimated, and through the proposed spatial filtering process, the accuracy of detected changes can be further improved. In this study, the proposed method of uni-temporal image change detection was applied to detect oil slicks on the ocean surface. According to the analyzed results, the false alarms of detected oil slicks can be less than 10%. Furthermore, the false alarms of detected oil slicks can be improved to less than 5% if a spatial filtering process is further applied.
Multi-temporal image change detection is mainly used to detect the differences between images from different dates and extract the corresponding temporal changes. In general, change detection of multi-temporal images requires two essential steps: (1) normalization of images acquired on different dates and (2) detection of changed areas from normalized images. Images must be normalized because images from different dates can have significant radiometric differences caused by varying environmental factors during image acquisitions, specifically factors related to atmospheric effects. Theoretically, the radiometric calibration procedures can be used to adjust these differences, but the process is usually costly, labor-intensive, and unfeasible for frequent periods and large area operations. For change detection purposes, relative image normalization is a better choice because instead of calibrating the images of different dates to reflectance level, the images are relatively normalized to share the same reference level, which usually is sufficient to reveal the real temporal changes.
The most common approaches for image change detection usually compare images from different dates and derive some measurements to quantify the changes
thresholds are then set to extract the changed area. Nevertheless, the optimal thresholds can vary from case to case and remain a challenging issue. In this study, image normalizations were carried out by using pseudo invariant features (PIFs). Once the PIFs are extracted from source images, they can be used to perform the normalization process. For image change detection, we adopted a method from the concept of PIFs extraction to obtain a set of pseudo variant features (PVFs) corresponding to changed pixels. Once PVFs are found, they can be applied as a reference to detect the changes from the spectral differences derived by change vector analysis. The experimental results indicate that the proposed method can offer quality PVFs as a reference for detecting changes from images acquired on different dates. According to the experimental results with various image sets, the accuracies of change detection are around 90% with 0.8 kappa coefficients.
The major differences between proposed uni-temporal and multi-temporal image change detections are the procedures of deriving change measurement. In this study, in uni-temporal image change detection, the degree of anomaly was used as change measurement and obtained based on the statistics of normal image background. In multi-temporal image change detection, the change measurement is carried out by the spectral differences obtained from normalized multi-temporal images. However, the method for obtaining references for changed and unchanged objects should be the most important component of change detection algorithms. In this study, methods to find references for changed and unchanged objects were proposed for both uni-temporal and multi-temporal image change detections. In practice, these methods for obtaining references can not only be used in the change detection processes of this study, but can also be applied to other change detection algorithms that require references for changed or unchanged objects. |
