運用深度學習框架結合遙測資料進行崩塌地變遷偵測與分類;Landslide Change Detection and Classification Using Deep Learning Frameworks and Remote Sensing Data

NCU Institutional Repository > 工學院 > 土木工程研究所 > 博碩士論文 > Item 987654321/97319

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/97319

题名:	運用深度學習框架結合遙測資料進行崩塌地變遷偵測與分類;Landslide Change Detection and Classification Using Deep Learning Frameworks and Remote Sensing Data
作者:	蔡尚年;Tsai, Shang-Nien
贡献者:	土木工程學系
关键词:	崩塌地;深度學習;語意分割;變遷偵測;二階段分析框架;Landslide;deep learning;semantic segmentation;change detection;two-step analysis framework
日期:	2025-07-28
上传时间:	2025-10-17 11:08:04 (UTC+8)
出版者:	國立中央大學
摘要:	山崩具有強大的破壞力與高度不可預測性，使其成為最嚴重的自然災害之一，嚴重威脅人類的生命與財產安全。台灣位於板塊交界處，且受低緯度熱帶與副熱帶氣候影響，地震頻繁、地形複雜，並伴隨豐沛降雨與颱風侵襲，使台灣成為山崩災害的高風險區域。因此，如何快速且精確地偵測山崩範圍，協助政府進行災害監測與防範，具有重要的實務價值。本研究提出一種結合深度學習語意分割技術的二階段分析框架，用於崩塌地變遷偵測與分類。該框架設計考量三個層面：(1) 資料面：輸入特徵涵蓋光譜、植生、紋理與地形等多元資料，具高度異質性，透過分階段訓練可降低學習難度；(2) 任務面：多類別分類具挑戰性且樣本不均，先進行變遷偵測再分類，有助模型聚焦變異區域；(3) 架構面：所用模型雖專長於像素級分類，卻不具備變遷特徵提取模組，藉由第一階段產製變遷圖，可有效引導後續分類。整體流程整合雙時相影像、單時期地形因子與變遷特徵，並透過逐步訓練提升模型效能。第一階段以未正規化特徵進行二元分類，區分變遷與非變遷區域；第二階段則使用正規化特徵並結合第一階段結果，將變遷區域細分為植生與裸露地、舊崩塌地、新崩塌地與植生恢復區等四類。此逐步訓練策略能有效提升模型的分類效能，特別是在變遷細節辨識上。本研究以2016 年與 2022 年的荖濃溪流域，以及 2019 年與 2021 年的台灣中部山區進行模型訓練與測試。模型表現以精確度（Precision）、召回率（Recall）、F1 分數（F1-score）和整體準確度（Overall Accuracy）進行評估。研究結果顯示，二階段分析策略能有效提升分類精度，具體而言，模型在兩個研究區的舊崩塌地類別中 F1-score 超過 0.9，在新崩塌地類別中超過 0.7，在植生恢復類別中超過 0.7 並達到 0.8。這些結果證明所提出的框架在崩塌地變遷偵測與分類中具備良好效能，對未來災害監測與管理應用具有潛在的幫助。 ;Landslides possess immense destructive power and high unpredictability, making them one of the most threatening natural disasters causing significant threats to human lives and property. Taiwan is located at the junction of tectonic plates and is influenced by a tropical and subtropical climate due to its low-latitude location, leading to frequent earthquakes, complex terrain, and extreme weather. These factors make Taiwan highly vulnerable to landslide disasters. Therefore, rapidly and accurately detecting landslide occurrences is crucial for disaster monitoring and management. This study proposes a two-step analysis framework that integrates deep learning with semantic segmentation techniques for landslide change detection and classification. The framework is designed based on three key considerations: (1) Data perspective: the input features consist of diverse and heterogeneous data sources, including spectral, vegetation, textural, and topographic factors. A stepwise training strategy is adopted to reduce learning difficulty; (2) Task perspective: multi-class classification of landslide changes is challenging due to sample imbalance. Performing change detection before classification helps the model focus on areas of interest; (3) Structural perspective: although the adopted model specializes in pixel-wise classification, it lacks a dedicated module for extracting change features. Thus, a preliminary change map generated in the first step effectively guides the subsequent classification. The overall process integrates bi-temporal imagery, single-period topographic factors, and derived change features, with progressive training to improve model performance. The two-step strategy begins with the model learning raw, unnormalized features to perform binary classification, distinguishing changed and unchanged regions. In the second step, the model utilizes normalized features combined with predictions from the first step, further refining the classification into four categories: vegetation and bare land, old landslide, new landslide, and vegetation reclaimed. This progressive training strategy effectively improves classification performance, particularly in detecting subtle change details. Two study areas were selected for model training and testing including the Laonong River Basin (2016 and 2022) and the Central Taiwan Mountain Region (2019 and 2021). Model performance is assessed using Precision, Recall, F1-score, and Overall Accuracy as evaluation metrics. Results of this analyze demonstrate that the two-step analysis strategy effectively improves classification accuracy. Specifically, the model achieved more than 0.9 F1-scores for old landslide in both study areas, over 0.7 for new landslide, and exceeding 0.7 and reaching 0.8 for vegetation reclaimed. The effectiveness of the proposed framework in landslide change detection and classification has a great potential for improving future disaster monitoring and management.
显示于类别:	[土木工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	13	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....