| 摘要: | 光學影像基於其高分辨率的光譜特性，經常被應用於進行崩塌偵測，但光學影像在應用上的主要限制為受雲層覆蓋時無法有效地、即時地進行崩塌偵測。合成孔徑雷達（SAR）的訊號可以穿透雲層覆蓋並透過檢測迴波特性來識別地表的變化，因此應用雷達影像在災害的緊急監測上擁有更高的策略彈性，有助於設計、執行與評估災害風險和災後復原計劃。本研究針對事件型崩塌(如颱風、地震等災害事件誘發之崩他)，應用變遷偵測技術，嘗試整合遙測光學影像與雷達影像在崩塌監測中的優勢，發展一並基於物件導向分析法（OBIA）之事件型崩塌偵測方法，使其能夠在多雲霧的天氣條件下保持高精度的崩塌偵測成果，以提高救災效果，減少災害發生期間和之後的人員傷亡與財產損失。
本研究選取台灣南部的荖濃溪流域作為研究區，該地區在2009年莫拉克颱風期間因暴雨而引發了大規模山崩並造成嚴重的破壞。具體分別使用事件前與事件後的FORMOSAT-2光學影像及L波段的ALOS-PALSAR雷達影像計算兩種影像指數：標準化植被指數差(NDVIdiff)和標準化Sigma指數(NDSI)。使用物件導向分析法，產製六張光學與六張雷達紋理圖像（平均值，標準差，對比，熵，同質性和相異性）進行崩塌探測。因為在山區中光學與雷達影像有著座標對位的問題而難以實現像素級和特徵級影像融合，因此本研究採用決策級融合法。考慮到近年來機器學習演算法在圖像分類中的成功應用，本研究使用了人工類神經網絡(ANN)分類器。總體進行了五項試驗：（1）光學影像試驗(ANN-NDVIdiff)、（2）雷達影像試驗(ANN-NDSI)、（3）光學及雷達影像整合試驗(ANN-All)、（4）光學影像受雲霧影響試驗以及(ANN-Cloud)（5）以雷達影像補償光學影像中之雲霧試驗(ANN-Mosaic)。
實驗結果表明，由光學及雷達影像整合試驗成果(ANN-all)取得之崩塌偵測結果比其他試驗結果有更高的總體精度和Kappa係數，顯示了光學與雷達影像的融合成果在崩塌偵測任務中的可行性與適用性。此外，研究成果顯示僅使用雷達影像之試驗(ANN-NDSI)仍可以探測到崩塌區域的大致位置，尤其適用於大規模崩塌。本研究認為在惡劣天氣情況下，雷達影像可以代替光學影像進行緊急的崩塌偵測作業。
;Optical remote sensing data has been used to assist the preparation of landslide inventory based on its high distinguishability of spectrum characteristic. However, its applicability can be limited when the applied image is contaminated by cloud. Synthetic Aperture Radar (SAR) signals can penetrate cloud and identify land surface changes by examining their backscattering characteristics which are useful information for landslide detection. Therefore, SAR-based emergency mapping has recognizable potential to design, implement, and evaluate disaster risks and recovery programs. The major objective in this study is to develop a method that can integrate both optical and SAR remote sensing data for an effective and rapid landslide detection, and aims to maintain high accuracy during cloudy weather conditions.
Accordingly, to integrate advantages of both remote sensing data in landslide detection, decision-level image fusion approach in an object-based image analysis (OBIA) framework is designed and tested in this study. The Laonong River watershed in southern Taiwan is selected as the study site, where heavy rainfall induced large-scale landslides and caused severe damages during Typhoon Morakot in 2009. Specifically, two indices, Normalized Difference Vegetation Index difference (NDVIdiff) for optical data and Normalized Difference Sigma-naught Index (NDSI) for SAR data, were separately calculated by using pre- and post-event Formosat-2 images and L-band ALOS-PALSAR images respectively, and then were applied in the OBIA to generate six texture indices (mean, standard deviation, contrast, entropy, homogeneity and dissimilarity) for both optical and SAR data. Decision-level fusion is adopted in this study when pixel-level and feature-level fusion are difficult to practice due to co-registration issue between optical and SAR images over mountainous area. In the study, considering the successful application of machine-learning algorithm in image classification in recent years, Artificial Neural Network (ANN) classifier is therefore applied. Overall, five experiments were performed and simulated: (1) ANN-NDVIdiff, (2) ANN-NDSI, (3) ANN-All, (4) ANN-Cloud and (5) ANN-Mosaic.
Finally, landslide detection results obtained by ANN-All show higher overall accuracy and Kappa coefficient than other experiments, indicating the significance of applicability of optical-SAR fused data in landslide mapping task. In other hand, result of ANN-NDSI indicates that SAR image can detect landslide areas in approximately location, especially performs well for large scale landslides. SAR image can instead of optical image during cloudy weather situation for emergency landslide mapping. |
