陳彥安 數位影像相關係數法應用於衛星影像崩塌地判釋之研究
Application of Digital Image Correlation Method to Identify the Landsliding based on Satellite Image 摘要 由於衛星影像具有資料獲取週期短、可迅速掌握地表改變狀況且涵蓋面積廣、資 料量大、具多光譜與重覆觀測之能力等特性，因此利用遙測技術來得知災害發生 地點與數量，已逐漸取代因天氣、資金、路況受限的現場勘查之傳統方法。 本研究利用風災前後的多頻譜衛星影像，使用數位影像相關係數法(Digital Image Correlation，DIC)計算兩張影像在可見光波段上數值資料的相關性，以設定的門 檻值來進行風災後新增崩塌地的判釋，依據分析結果探討數位影像相關係數法在 新增崩塌地的判釋效果。 數位影像相關係數法原本是尋找兩影像間之最高相關性，本研究中兩張崩塌前 後的影像中地貌改變的區域會造成多頻譜波段之數值差異，因此將數位影像相關 係數法稍做修改，以尋找兩影像間相關性較低的位置，來判釋崩塌的發生。在 DIC 法判釋結果中會有誤判的情形發生，本研究利用常態化差異植生指數 (Normalized Difference Vegetation Index，NDVI)和數值高程模型(Digital Elevation Model，DEM)之特性，來降低誤判之結果。 本研究 DIC 法由大網格 16*16 門檻值 0.997 初步分析後再進行小網格 5*5 門檻 值0.998 的 DIC 分析結果，正確率可達 76.58%，代表大部分新增加的崩塌地都能 判釋的出來。為降低誤判率，本研究採用NDVI 來過濾掉崩塌前裸露區域和崩塌 後植生復育區域及以DEM 過濾掉河川旁和村落這些坡度平緩的區域之誤判，整 體而言DIC 法在判釋崩塌前後兩影像中地貌的差異有良好的判釋效果。
The satellite image has following characteristics: short capture period, ability to detect the change of the earth surface rapidly, large coverage, large amount of information, multiple spectra and the ability of repeating observations. Therefore, instead of the traditional method of site investigation, which is limited by weather conditions, expenditure, and traffic situation, the remote sensing technology based on satellite image is gradually used to determine the occurrence locations of landslide disasters and their quantities.
In this study, we apply the digital image correlation method to evaluate the correlation of two multi-spectral satellite images before and after disasters. A threshold value is used to determine whether the new landslides occur after the disasters. The effect of applying DIC method to identify the landslide is then discussed.
However, the change of the earth surface will induce a difference between
multi-spectral signal of the satellite images before and after the landslide. Therefore, the DIC method can be easily modified to determine the occurrence of landslides. The position with lower correlation between two images will be found out. There exist also many misjudgment situations in the result We use the characteristics of the Normalized Difference Vegetation Index (NDVI) and the Digital Elevation Model (DEM) to reduce the rate of misjudgment.
In this study, the analysis procedure is divided into two steps. The whole area is roughly analyzed with a larger subset at first. Then, the area, which is judged to be a landslide area in the rough analysis, is then closely analyzed with a small subset. The result shows that the accuracy rate can reach 76.58% in the analysis carried out with a big subset 16*16 pixels and threshold value 0.997 in the rough analysis and a small subset 5*5 pixels and threshold value 0.998 in the close analysis. This means that a lot of new occurrence landslide is identified. For reducing the rate of misjudgment, we use NDVI to filter the misjudgment region, which is uncovered before the disaster and vegetative restorative after the disaster. DEM is also applied to filter the
misjudgment region, which lies around the smooth riverbed and village. This study shows that DIC method can be applied to identify the landslides based on the satellite images before and after the landslide.
