Please use this identifier to cite or link to this item:
標題: 相對輻射糾正法應用於SPOT衛星影像變遷偵測影響之評估-以台中縣烏石坑地區為例
Evaluation of the Effect of Relative Radiometric Normalization on Multi-date SPOT Image Change Detection
作者: 潘麗慧
Pan, Li-Hui
關鍵字: 直方圖匹配
Histogram Matching
Pseudo-invariant Features
Linear Regression Normalization
Image Regression
Normalized Difference Vegetation Index
Root-mean-square Error
出版社: 森林學系
摘要: 多期影像變遷偵測通常應用於地景變遷監測,但因非地表因子如大氣效應、光照、觀視幾何和感測儀校準及地表因子季節物候和地形效應等之變異形成假變遷,以致於降低變遷偵測之精度。本研究探討非地表因子及植物物候形成影像光譜變異對變遷偵測之影響,並確認影像同態化對變遷偵測之重要性。同時評估影像同態化法對變遷偵測之成效,找出適當之方法可提高辨識率。直方圖匹配、擬似不變異地物線性迴歸及全影像迴歸等三方法被應用在台中縣烏石坑地區之十三期SPOT影像上,並評估其同態化效果。就影像外觀、均方根誤差、頻度分布、變遷偵測試驗及操作難易來看,直方圖匹配在同態化之效果最優,全影像迴歸其次,而擬似不變異地物線性迴歸之效果最遜。就波段來看,兩可見光波段之同態化效果優於NDVI,且紅光略優於綠光,近紅外光之效果最差。變遷偵測率降低之主因在無變遷誤授為疑似變遷或有變遷,HM法及IR法可有效改善,而PIF法則起伏不定。另一方面,有變遷之漏授雖佔整幅影像甚小比例卻是變遷偵測之關鍵。各方法之改善漏授效果缺乏規律性,似因樣本不足之故,仍需深入研究。同態化效果有傾向隨著從屬參考影像對之季差縮短而增加,因此像對最好同屬夏季,同屬秋春次之,避免分屬冬夏。因此,影像同態化整體而言對變遷偵測是相當重要的。
Change detection from multi-temporal images has been commonly used in monitoring changes in landscape. However, spurious changes caused by variations of non-surface factors including illumination, viewing geometry, sensor calibration, and atmospheric effects, as well as surface factors including seasonal phonological differences and topographic effects, may lead to inaccurate results, thereby reducing the accuracy of landscape change detection. The study was intended to examine the influence of spectral variations caused by the non-surface factors and phonology on change detection and to confirm the importance of image normalization to change detection. The study also evaluated the performance of three image normalization methods in relation to change detection. Histogram matching (HM), image regression (IR), and pseudo-invariant feature (PIF) regression requiring the use of a reference-subject image pair, were applied to thirteen-date SPOT images from the Wu-Shyr-Keng area of Taichung Prefecture. Three images out of thirteen, acquired in summer, winter, and spring (or fall) were selected as a reference image, respectively. The methods were compared in terms of their capability to improve visual image quality, statistical robustness, and ease of implementation. The HM method was the first in overall performance, IR was the second, and PIF was the last. Image normalization had better performance using visible bands than using NDVI and near-infrared bands, and using red band slightly better than using green band. Low accuracies in change detection were primarily due to the erroneous assignment of no-change pixels to the likely-change class or true-change class. HM and IR improved accuracies effectively, but PIF performed unsteadily. On the other hand, the omission of true-change pixels is the key to change detection in spite of usually occupying a small proportion of the entire image. The three methods performed unsteadily, probably due to insufficient sample of true-change pixels, and further research shound be conducted in the future. The performance also tended to increase with a reduction in season difference between an image pair. An image pair acquired in the same season, particularly in summer, with time difference as short as possible is preferred, but one image of an image pair acquired in winter with the other in summer, or vice versa, should be avoided. Thus image normalization overall is of great importance for change detection.
Appears in Collections:森林學系



Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.