Landslide Mapping at the Watershed Level with OBIA — Case studies in the Baichi watershed, Taiwan

崩塌製圖是一個進行崩塌危險評估的關鍵步驟。本文運用物件導向影像處理方法於台灣白石集水區兩個個案的崩塌檢測與製圖。在第一個個案中發展出一個多尺度的物件導向影像處理方法來繪製2004年艾莉颱風後的崩塌，檢測的準確率達到86%，此作法另可檢測出各種尺寸的崩塌。第二個個案所用的是地理物件導向影像變化處理方法(GEOBICA)繪製多重時間的崩塌。以此方法檢測2004年以及2008年的崩塌，兩者檢測結果均達到83%以上的準確率。此作法更可提供四年之間崩塌的發生、回復、持續作用的線索。本論文並分析此方法的限制，提出後續研究的方向及可能性，提供學界作為參考。Landslide mapping is a fundamental step to undertake landslide hazard and risk assessments. We designed two case studies in the Baichi watershed, Taiwan, to detect and map landslides with object-based image analysis (OBIA) techniques. We first developed a multiscale OBIA approach to map shallow landslides after Typhoon Aere in 2004. The detection performance reached 86.5% in training and 86% at validation. The method detected substantial number of landslides of all sizes. It could serve for mapping shallow landslides over large areas. We developed a second landslide detection scheme, based on a geographic object-based image change analysis (GEOBICA) methodology, to map landslide for multi-temporal inventories. The technique detected landslide at two dates, 2004 and 2008. Its detection performance was above 83% in both the training and validation watersheds. The technique also provided an account of landslide recovery, occurrence and persistence in the watersheds over the 4-year period. This methodology could serve to develop multi-temporal inventories, to check old landslide maps, and to identify landslide hotspots. Both techniques presented inherent limitations: site and image specificity of the classification thresholds in the first case, a restricted geographic scope and the necessary absence of shadow and clouds in the second case. We suggested further research directions in OBIA landslide mapping, to remedy current limitations and develop the potential of both methodologies. We finally recommended the use of OBIA in landslide hazard and risk assessment, beyond its current landslide mapping scope