近年來，臺灣山區的開發與擴張使得邊坡安全評估變得更加重要，特別是在豪雨事件後更容易發生山崩，政府長期以來將防災和減災列為重要的施政項目，並且在地質法通過後，臺灣都市周邊坡地的山崩潛勢評估和地質敏感區的劃設成為關注的重點，因此繪製台灣全島山崩目錄是極為重要的，在深度學習方法尚未盛行前，前人使用完全人工的方式對遙測影像中的山崩部分進行一一圈繪，此方法需耗費非常大量的時間以及人力。隨著時代、科技的演進，深度學習方法在許多領域獲得巨大的成功，利用深度學習我們可以電腦軟體先行圈繪大部分之山崩區域，再經由專業人士進行評估，以此達到減時之效果。 遙測影像具有高空間解析度和豐富的地物資訊，但由於台灣山體地形複雜、環境變化大，傳統的語意分割方法難以達到理想效果。為了解決這一問題，本論文以UNetFormer為主要架構下，提出了一種結合UNetFormer、對比式學習及GAN鑑別器的混合模型，旨在提升分割結果的真實性、邊界細節和模型的泛化能力。 經過實驗後，我們提出的架構與現今流行的遙測影像語意分割模型以及與原架構UNetFormer比較過後，我們取得了不錯的成績。我們提出的兩大組件對比式學習以及GAN鑑別器經過消融實驗的驗證後，確實能有效提升模型的分割能力。 ;In recent years, as Taiwan′s mountainous regions have expanded, evaluating slope safety has become crucial, especially after heavy rains which increase landslide risks. Following the Geology Act, assessing landslide potential near urban areas is now a focus. Creating a comprehensive landslide catalog for Taiwan is essential. Before deep learning, manual annotation of landslides in remote sensing images was time-consuming and labor-intensive. With technological advancements, deep learning has greatly improved efficiency in many fields. Using software for preliminary annotations, followed by expert review, saves significant time. This thesis presents a hybrid model using UNetFormer architecture, incorporating contrastive learning and a GAN discriminator, designed to improve the effectiveness of semantic segmentation in complex terrains. Our multi-stage training framework enhances segmentation accuracy, boundary precision, and model adaptability. Experimental results show our model′s superiority in segmentation capabilities compared to traditional methods and the original UNetFormer.
