| 摘要: | 崩塌是台灣山區一種非常常見的自然災害，對於崩塌災害管理，分析崩塌之特徵、發生源及過去的災害事件非常的重要，而在崩塌災害分析中，崩塌資料庫擁有高度需求，因此資料庫的建置是一項基本且重要的工作，崩塌資料庫繪製的工作在遙測技術及影像分析的進步下變得非常有效率，本研究利用深度學習架構並使用衛星遙測資料進行崩塌地識別以建構崩塌資料庫。而進行崩塌地識別時需要有足夠且適當的資料且此些資料須與崩塌地有高度相關，因此，本研究收集及數化由2000至2014發生於石門水庫集水區及荖濃溪流域的崩塌地圖資作為訓練之標記資料。另外，本研究使用了光學衛星影像、常態化差值植生指標(NDVI)、灰度共生矩陣(GLCM)紋理特徵及地形因子作為訓練之輸入資料。其中，前三者為雙時相資料，而地形因子僅包含單一時相。而因為研究中訓練資料集中的輸入資料包含不同時序的二維影像，並有多樣化特性，因此，此研究建置一「兩步驟」的訓練架構以因應輸入資料之不同特性；第一步驟中將特徵擷取模型ConvLSTM U-Net用於萃取雙時相影像中的特徵，而在第二步中將隨機森林(Random Forest)模型用於分析特徵圖及地形因子以產生最後崩塌地辨識結果。;Landslide is one of the common natural hazards in the mountainous regions of Taiwan. For landslide management, it is very important to analyze the characteristics, occurrence and past-hazard events of landslides. Therefore, landslide detection is a fundamental and important task, and it is also in high demand for disaster assessment. Combining with remotely sensed data and image analysis landslide mapping task is time-effective nowadays. This study conducts landslide detection using deep learning framework with satellite remote sensing images and spatial data. Landslide recognition requires adequate data with high relationship to the occurrence of landslides. Therefore, this study manually collects past hazard events from 2000 to 2014 in Shimen and Laonong watersheds, Taiwan as training dataset. In addition, available features including optical satellite images, NDVI, gray level co-occurrence matrix (GLCM) texture features and topographic factors are adopted for detecting landslide. The former three components are bitemporal images, yet topographic factors only contain one-time stamp. Because the training data-used contains time series of two-dimensional images and multi-types input features, this study integrates feature extraction and classification processes into a two-steps training framework to adapt to the existing data for landslide identification. In the first step, ConvLSTM U-Net model is used as feature extraction model for extracting features from bitemporal images. In the second step, random forest is utilized as a final classifier to analyze input data of feature maps and topographic elements.
The trained model is employed to create landslide maps in selected ten test regions. The detected landslide maps are evaluated by quantitative indexes, e.g., Precision, Recall, F1-score, IOU and Kappa. The assessment results indicate that the developed deep learning framework can achieve the accuracies of higher than 0.7 of IOU, 0.8 of F1-score and 0.8 of Kappa. |
