多期衛星影像相對輻射糾正於台中縣烏石坑地區變遷偵測之探討

Radiometric Normalization of Multi-date Satellite Images for Change Detection in the Wu-Shyr-Keng Area of Taichung Prefecture

多期影像變遷偵測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形成假變遷，以致變遷偵測正確率降低。本研究探討非地表因子及物候導致影像光譜變異對變遷偵測之影響，確認影像同態化對變遷偵測之重要性，同時評估各影像同態化法之功效。本研究將擬似不變異地物線性迴歸（PIFLR）、全影像迴歸（IR）、直方圖匹配（HM）等三方法應用在台中縣烏石坑地區之十三期SPOT影像上，以評估其同態化效果。就影像外觀、操作難易、均方根誤差及變遷偵測正確率來進行比較，HM之效果最優，IR居次，而PIFLR最遜。就波段來看，兩可見光波段之效果優於規整差植生指標（NDVI），且紅光略優於綠光，而近紅外光最差。變遷偵測正確率降低之主因在無變遷誤授為疑似或有變遷，HM及IR法能有效改善，PIFLR法則起伏不定。另一方面，有變遷之漏授雖佔甚小比例卻是關鍵。各方法之改善漏授效果缺乏規律性，似因缺乏不同試區且具充足之真變遷樣本，仍需深入研究。然影像同態化對變遷偵測至為重要。 Change detection from multi-date images has been commonly used in monitoring changesin landscape. However, spurious changes caused by variations of non-surface factors includingillumination, 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 theinfluence of spectral variations caused by the non-surface factors and phonology on change detection andto confirm the importance of image normalization to change detection. The study also evaluated theperformance of three image normalization methods in relation to change detection. Pseudo-invariantfeature linear regression (PIFLR), image regression (IR) and histogram matching (HM) requiring the useof a reference-subject image pair, were applied to thirteen-date SPOT images from the Wu-Shyr-Kengarea of Taichung Prefecture. The methods were compared in terms of their capability to improve visualimage quality, statistical robustness, and ease of implementation. The result showed that HM method wasthe first in overall performance, IR was the second, and PIFLR was the last. Image normalization hadbetter performance by using visible bands than using normalized difference vegetation index, using redband slightly better than using green band, and the near-infrared band had the least. Low accuracies inchange detection were primarily due to the erroneous assignment of no-change pixels to the likely-changeclass or true-change class. HM and IR improved accuracies effectively, but PIFLR performedunsteadily. On the other hand, the omission of true-change pixels is the key to change detection in spite ofusually occupying a small proportion of the entire image. The three methods performed unsteadily,probably due to a scarcity of different test areas with sufficient true-change samples. Further researchshould be conducted in the future. Nevertheless, image normalization overall is very important forchange detection.

Change detection from multi-date images has been commonly used in monitoring changesin landscape. However, spurious changes caused by variations of non-surface factors includingillumination, 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 theinfluence of spectral variations caused by the non-surface factors and phonology on change detection andto confirm the importance of image normalization to change detection. The study also evaluated theperformance of three image normalization methods in relation to change detection. Pseudo-invariantfeature linear regression (PIFLR), image regression (IR) and histogram matching (HM) requiring the useof a reference-subject image pair, were applied to thirteen-date SPOT images from the Wu-Shyr-Kengarea of Taichung Prefecture. The methods were compared in terms of their capability to improve visualimage quality, statistical robustness, and ease of implementation. The result showed that HM method wasthe first in overall performance, IR was the second, and PIFLR was the last. Image normalization hadbetter performance by using visible bands than using normalized difference vegetation index, using redband slightly better than using green band, and the near-infrared band had the least. Low accuracies inchange detection were primarily due to the erroneous assignment of no-change pixels to the likely-changeclass or true-change class. HM and IR improved accuracies effectively, but PIFLR performedunsteadily. On the other hand, the omission of true-change pixels is the key to change detection in spite ofusually occupying a small proportion of the entire image. The three methods performed unsteadily,probably due to a scarcity of different test areas with sufficient true-change samples. Further researchshould be conducted in the future. Nevertheless, image normalization overall is very important forchange detection.