事件誘發山崩目錄為近年山崩研究之趨勢。現行崩塌地區調查方法中，衛星影像具不受時間、空間限制的優點，可於短時間獲得大區域面積調查結果，且取像頻率高，故適於利用多時期影像差異判釋誘發山崩。前人利用衛星影像進行自動判釋，多僅為單期影像之判釋，有以多時期影像結合變異向量分析法進行變遷偵測者，但亦僅利用變異向量大小與光譜角為分析依據。因為影像光譜角僅是前後期影像向量的夾角大小，所能代表的變遷資訊有限，所以本研究採用變異向量大小及位態，包含更完整資訊，作為自動判釋之基礎。 本研究利用艾利颱風前後時期於石門水庫集水區之SPOT 5衛星影像進行事件誘發山崩自動判釋。結合交互相關與趨勢面分析方法，使前後時期影像的對位精度達亞像元，並經直方圖匹配令多時期影像輻照度同態化，再利用變異向量分析法進行變遷偵測以獲得誘發山崩目錄。經試誤法與訓練樣本輔助，設定變遷門檻值為變異向量大小大於30 DN及變異向量位態之邊界條件，篩選艾利颱風誘發之可能山崩。最後，輔以坡度、原始影像光譜值及沖積層資訊去除非誘發山崩之地物變遷型態，完成艾利颱風事件誘發山崩目錄。分類誤差矩陣分析結果顯示，山崩組使用者準確度為78.20%，山崩組生產者準確度為96.25%，總體準確度達99.88%，Kappa指標亦達0.99，成果極為良好。;Construction of an event landslide inventory is a trend in recent landslide research. Using remote sensing images to construct a landslide inventory is time-saving and more thorough in an area. Researchers can get mapping results of wide areas in a short time. With high acquiring frequency, satellite images are suitable for interpretation of event-induced landslides by differences between multi-temporal images. Most previous studies used a single image to interpret landslides. Some else used multi-temporal satellite images and change vector analysis to detect. However, the latter only used change vector distance and spectral angle, the information is limited. The present study uses both change vector magnitude and orientation to detect landslides. This then contain much more information for the automated interpretation. This study used SPOT 5 satellite images before Typhoon Aere and after Typhoon Aere at the Shimen Reservoir catchment area. Two rectified satellite image was geometrically correlated and corrected to a sub-pixel level. And then, histogram matching made the irradiance of multiple images unifying. At last, change vector analysis was utilized for the detection of event-induced landslides. With try and error and aid of training sample sets, change thresholds were determined as change vector distance greater than 30 digital number (DN) and a specific change vector orientation space to obtain hot spots of possible landslides induced by the Typhoon Aere. These hot spots were finally screened by using slopes, alluvium, and satellite images, and to complete event landslide inventory of Typhoon Aere. The classification error matrix shows that user’s accuracy of landslide is 78.20%, producer’s accuracy of landslide is 96.25%, overall accuracy is 99.88%, and Kappa index also reaches 0.99. The study result is very good.
