| 摘要: | 全世界大部分國家都曾經遭遇崩塌地引發之後果，其形式包括經濟衰退、數十億美元的損失和更多的傷亡。崩塌地成因有先天性的自然因素或裸露的陡峭斜坡沖刷而形成，或者是由自然災害引起的如地震，火山爆發，積雪融化和豪雨等。氣候變遷是一個全球性挑戰，其極端氣候的暴雨，對有些易引發大規模崩塌地區，將面臨災難性後果。因此這些挑戰急需發展一套快速有效的評估系統，能產生錯誤率較低之崩塌分類成果，有效且精確判釋出崩塌區位與非崩塌區位供後續坡地防災應用。 台灣在1999年9月21日發生芮氏7.3級地震，造成灣中部地區山坡地大規模的崩塌裸露，對該地區之經濟、聚落、設施和生態系統造成嚴重破壞。本研究以受災較嚴重之九九峰地區為試區，透過資料探勘技術對震災後SPOT影像資料進行判釋分析，探討崩塌區位判釋之最佳方法。本研究評估九種資料探勘方法，首先以兩個光譜指數NDVI及TBI來萃取崩塌與非崩塌區位之光譜影像值，並建立其判釋模式萃取崩塌區位，並導入九種資料探勘方法，包括四個非監督分類法如K-means、Minibatch K-means、BIRCH和GMM等，以及5個監督分類法如SVM、DT、ET、RF和XGBoost等，進行崩塌分類成果及精度比較，並計算出各種定量統計數據以供驗證，包括總體精確度(OA)、Kappa值、使用者精度(UA)及生產者精度(PA)等。 研究結果指出K-means方法優於其他分類方法，其總體精確度為95.94%，相較於XGBoost為95.49％、DT和RF為95.38％，而SVM方法之總體精確度最低為92.51％。整體而言，這九種方法都有還不錯的表現，總體精確度皆達到90％以上，表示其在判釋崩塌地上具重要意義。而未來對於崩塌區位之判釋應用上，建議可結合類神經之深度學習演算法或加入其他地形因子以有效提升判釋之精確度。
Major countries worldwide face the consequences of landslides in the form of reduced economy, damages worth billions, and higher fatalities. Landslides are the result of upsetting steeper inclines that were previously preconditioned or bare of vegetation. They are caused by natural hazards such as earthquakes, volcanic eruptions, melting of snow, and heavy rainfall showers. Climate change represents a global challenge that induces frequent volcanic and seismic activities due to tectonic excitation, and torrential rainfall with an abnormal downpour. Some regions are not preconditioned to tolerate such extreme weather changes and face cataclysmic repercussions in the form of landslides. Such challenges call for urgent development of a fast and efficient assessment system generating error-free Landslide Inventory Maps (LIM) that depicts a clear boundary between affected and unaffected regions. Taiwan is one such country that has been enduring such calamities throughout recent years. One particular catastrophic incident in the form of an earthquake having a magnitude of 7.3 in Richter scales which occurred on 21st September 1999, devastated the central part of Taiwan inflicting serious damages to its economy, human livelihood, infrastructure, and ecosystem. Mt Jou-Jou one of the severely impacted regions was adopted for this study. Analysis of Geo-Spatial data with data mining is cost-efficient and reduces dangerous fieldwork. This research explores the potential of nine data mining techniques along with a pixel-based image differencing on two spectral indices i.e., Normalized Difference Vegetation Index (NDVI) and Total Brightness Index (TBI) derived from multi-temporal SPOT satellite imagery for landslide detection. Landslide maps were generated and compared from four unsupervised i.e., K-means, Minibatch K-means, Balanced Iterative Reducing and Clustering using Hierarchies (BIRCH), and Gaussian Mixture Models (GMM), and five supervised learning algorithms which included Support Vector Machines (SVM), Decision Tree (DT), Extra Trees (ET), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). For comparison, a validation set was employed from which various quantitative statistics were computed such as Overall Accuracy (OA), Kappa Statistics (K), User's Accuracy (UA), and Producer's Accuracy (PA). The results suggested that the K-means algorithm outperformed other algorithms and showed the highest overall accuracy of 95.94% with a close follow-up from XGBoost (95.49%), DT, and RF (95.38%). The lowest accuracy was yielded by the SVM algorithm of 92.51%. In general, all the algorithms delivered outstanding performance and achieved overall accuracies well above 90% indicating their significance for identifying landslides. The overview of the conclusion in this research stated that landslide mapping using data mining algorithms and SPOT-derived spectral indices can provide essential surface information that can further act as an efficient tool for future landslide detection problems & research such as comparison with other deep learning algorithms and further addition of topographic, geologic, morphologic and lithologic information to the dataset. However, the proposed system in this research can be further used in regions with similar topographic and geologic nature. |
