應用多變數轉化偵測法於多光譜影像變遷偵測

Application of Multivariate Alteration Detection to Change Detection in Multi-spectral Imagery

洪志賢(Chih-Hsien Hung) ； 陳錕山(Kun-Shan Chen)

《航測及遙測學刊》 12卷1期 (2007/03) Pp. 37-47

https://doi.org/10.6574/JPRS.2007.12(1).3

摘要

影像差異法為最簡單的變遷偵測方式，係兩時期影像相減。使用多光譜影像進行影像差異法，為了結合所有波段的變遷資訊，而將差值影像做主軸轉換。傳統主軸轉換使用差值影像的共變異矩陣之主成份分析(Principal Component Analysis)。本研究使用多變數轉化偵測法(Multivariate Alteration Detection, MAD)，以典型相關分析(Canonical Correlations Analysis)為基礎，考慮兩時期影像間的交叉共變異矩陣，先行主軸轉換。本方法的特色在於線性轉換之不變性，可去除兩時期影像間大氣輻射的影響。再以轉換後影像，利用卡方統計檢定法(Chi-Square Test)，判斷變遷區域，使偵測結果能更反應真實地表覆蓋的變遷。經過由差值影像與MAD成分所得的變遷二位元影像，證明利用線性轉換的不變性，去除前、後期影像之間輻射強度的差別，可省掉相對輻射校正的前處理，還能排除因季節性引起的變遷。由模擬影像在不同信心水準下，雜訊比與整體精度的趨勢圖，雜訊比大小對於信心水準的選取沒有影響。而整體精度在訊雜比為10以下較差，所以使用多變數轉化偵測法若其兩時期影像之訊雜比在10以下便會對偵測的結果有較大的影響。

關鍵字

變遷偵測 ； 影像差異法 ； 主成份分析

並列摘要

When detecting changes in panchromatic images taken at different points in time, it is customary to analyze the difference between two images. Areas with little or no change have zero or low absolute values, and areas with large changes have large absolute values in the difference image. If image data gives more than two channels, it is difficult to visualize changes in all channels simultaneously. To solve this problem and to collect information on change, linear transformations of the image data can be considered. Traditionally, we make linear transformation by using principal component analysis via the covariance matrix of difference between two images. In this study, we perform linear transformation by applying multivariate alteration detection (MAD) by cross-matrix between two images. The property of the multivariate alteration detection transformation is the linear scale invariance. So, if we use MAD, preprocessing by linear radiometric normalization is superfluous. To detect the change areas by Chi-Square test, and the major changes is directly related to target changes, not seasonal or atmospheric effects. Results verify the effectiveness of the MAD method for change detection of multi spectral images.

並列關鍵字

change detection ； principal component analysis ； canonical correlation analysis