應用NDVI植生指標與平均值調整影像分割法於崩裸地萃取－以六龜試驗林地區多期福衛二號影像為例

Using the NDVI and Mean Shift Segmentation to Extract Landslide Areas in the Lioukuei Experimental Forest Region with Multi-temporal FORMOSAT-2 Images

謝漢欽(Han-Ching Hsieh) ； 鍾智昕(Chih-Hsin Chung) ； 黃俊元(Chun-Yuan Huang)

《台灣林業科學》 32卷3期 (2017/09) Pp. 203-222

https://doi.org/10.7075/TJFS.201709_32(3).0003

摘要

面對高解析度光學衛星影像用於地表目標物的分類，傳統上以像元為基礎的分類方法，往往效果不令人滿意；近年來採用影像分割技術，以物件為基礎的影像分類程序可以得到更好的結果。本研究以高雄市六龜試驗林及其周邊29,400公頃土地為試驗區，選用福衛二號2006年、2009年及2011年3期夏季影像，經影像輻射校正前期處理及不改變光譜特性的全色影像銳化融合處理，建立各期影像NDVI指標，可減除地形效應。應用具多項優點的平均值調整影像分割法及簡易物件形狀指標及坡度法則篩選程序，自動萃取3期試驗區的崩裸地。使用樣區系統取樣法，以各期樣區人工數化的崩裸地當作地真值，改進崩裸地自動萃取的精度評估。結果3期綜合分類準確度皆達95%以上，Kappa係數達85%以上，顯示本研究的影像分割程序對多期FS2影像的崩裸地的自動萃取具有好的結果，具可應用性。從3期崩裸地面積變遷分析結果可得知，2009年莫拉克颱風造成非崩裸地轉變成崩裸地的轉移率是後續兩年期間的3.5倍，而2009年至2011年間由崩裸地轉為非崩裸地的植生復育率是2006年至2009年的18.7倍。此外將2009年至2011年新增的崩裸地，透過數值地形衍生的向源河系套疊，可證明新增崩塌地多位於原有崩塌地往上游之向源侵蝕區域。本研究成果有利於六龜試驗林莫拉克颱風後崩塌地的空間治理。

關鍵字

影像分割 ； 崩塌地 ； 福衛二號 ； 植生指標

並列摘要

When applying high-spatial-resolution satellite imagery to classify of earth surface targets, traditional pixel-based classification methods often produce unsatisfactory results. In contrast, using object-based image classification with image segmentation approaches over the last decade achieved further improvements. In this study, 29,400 ha of forestland involving the Lioukuei Experimental Forest and the surrounding area in Kaohsiung City was used as the experimental area. Three Formosat-II (FS2) images of the experimental area obtained in the summers of 2006, 2009, and 2011 were respectively selected and preprocessed through radiometric calibration and pan-sharpening fusion without distorting their spectral characteristics. The normalized difference vegetation index (NDVI) of each FS2 image was established for follow-up image segmentation with a mean-shift algorithm. After NDVI segmentation with the mean-shift method, landslide and bare (LSAB) areas for each date were automatically extracted using a simple rule including a shape index and average slope of an object to filter out noise objects and mountain village areas. To improve the accuracy assessment for extracting of LSAB areas for each date, an assessment was performed by comparing to manually digitized sub-images of each pan-sharpened image through a system sampling method. Results showed that the overall accuracies of the 3 dates were ＞ 95%, and their Kappa coefficients were ＞ 85%; thus, the mean shift procedure can successfully be applied to extract LSAB areas from multi-date FS2 images. The area transition rate of Un-LSAB to LSAB areas between 2006 and 2009 after typhoon Morakot was 3.5-times that in the following 2 yr, and from 2009 to 2011, the area transition rate of vegetation regeneration was 18.7-times that between 2006 and 2009. Inspecting the spatial distribution of newly increased LSAB areas from 2009 to 2011 with extended tributaries derived from digital elevation model, transition areas from vegetation cover to LSAB were mostly located in headward areas.

並列關鍵字

NDVI ； mean shift segmentation ； FORMOSAT-2 ； landslide