請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92119
標題: | 應用機器學習方法於地層下陷分析模式建構與未來趨勢推估 Modeling Land Subsidence and Forecasting its Future Trends with Machine Learning Techniques |
作者: | 林培宇 Pei-Yu Lin |
指導教授: | 韓仁毓 Jen-Yu Han |
關鍵字: | 地層下陷預測,衛星定位系統,皮爾森空間自相關,時間序列分解,機器學習, Land Subsidence Forecasting,Global Positioning System (GPS),Autocorrelation of Pearson,Time Series Decomposition,Machine Learning, |
出版年 : | 2024 |
學位: | 碩士 |
摘要: | 地層下陷是一種土層垂直沉降之地質災害,由於經常發生在用水量較大之農業、住宅地區,過去常將其歸咎於超抽地下水之後果。地層下陷分析模式可分為統計經驗法、理論分析法與人工智慧方法。前兩者大多專注於探討地層下陷與地下水位變化的關聯性,然而統計經驗法需蒐集大量觀測資料以建立誤差小之分析模式,理論分析法則須透過現地觀測與室內壓密試驗估計水文地質參數以描述一地區之土壤透水性質與壓縮特性才得以有效模擬地層下陷行為。上述方法並不適用於缺乏觀測資料且水文地質參數未經估計之地區,且水文地質參數也因其空間異質性而僅適用於單一地區之地層下陷模擬。雖然機器學習、深度學習等人工智慧方法可以直接建構各影響因子與土層壓縮量之非線性關係,在處理具有趨勢性觀測資料之預測任務,如未來之地層下陷量預測,模型將因為測試資料與訓練資料的值域不同,而無法獲得良好的樣本外預測結果。因此,本研究欲建構一適用於複數地區之地層下陷分析模式以提供大範圍之地層下陷趨勢,預計可作為政府實施自然保育、用水管制政策之依據。本研究之目標為二:(一)建立地層下陷擬合模型,並探討各影響因素與地層下陷之間的作用機制,(二)利用現有觀測資料對地層下陷執行樣本外預測,並比較原始觀測資料與去趨勢性觀測資料之預測成果。本研究以空間自相關分析評估105至110年來主要地層下陷地區,並透過機器學習方法模擬雲林地區近年來之地表高程變化,模型之自變數可分為地下水位因子、土地使用因子與地質地形因子。研究成果顯示考量上述地層下陷影響因素後,本研究建構之地層下陷分析模式於單一GPS固定站之均方根誤差約為3公分,最低可至0.2公分;應用於推估未來地表下陷量的準確度可達1.4公分,且具備捕捉地表因地下水位變化造成之季節性波動的能力。此外,部分GPS固定站之土地使用因子對地表高程變化模擬之貢獻度甚至高於地下水位因子。綜合上述,本研究認為短期內的地下水位僅能描述季節性之地層下陷行為,必須加入其他輔助資料才能完整描述其空間異質性。並且以機器學習方法模擬地層下陷時,應根據不同土層的沉陷特性區分趨勢性與季節性,且分別建立兩個模型以提高未來地層下陷趨勢變化的預測準確度。 Land subsidence is a type of geohazard that causes vertical settlement of the earth’s surface. Often observed in agricultural and residential regions, land subsidence was attributed to the over-exploitation of groundwater. Theoretical approaches based on the relationship between ground surface level and groundwater level require field or laboratory experiments to obtain more accurate parameters to describe the hydrogeological properties of rocks. However, these approaches are not applicable in places where hydrogeological datasets are unknown. While AI methods can construct the non-linear relationship between land subsidence and its effective drivers, predicting trends like future subsidence using machine learning encounters challenges due to data discrepancies. Understanding the future trend of land subsidence in these places allows the government to imply appropriate mitigation strategies, thereby avoiding economic and environmental losses. The objective of this study is twofold: first, constructing a land subsidence prediction model that can be applied without prior knowledge of hydrogeological properties; second, forecasting the amount of land subsidence with existing observations. In this study, machine learning techniques were used to model the surface ground height in Yunlin, Taiwan from 2016 to 2020. Machine learning techniques can show a non-linear relationship between different datasets, including land-use change, lithology, and topological conditions. The results indicated over 99% of the variance in the ground surface level could be explained, and the root mean squared error of simulation was below 3 cm. Results also show that our model significantly improves predictions after combining time series decomposition, providing precise subsidence forecasting with 1.4cm accuracy. According to our simulation, land subsidence might be more strongly influenced by land use type than changes in groundwater level in some circumstances. These findings suggest that when simulating land subsidence, developing separate models for trend and seasonal components can enhance the accuracy of predicting future trends in land subsidence. Short-term groundwater level changes can only depict the seasonal behavior of land subsidence; auxiliary data should also be considered. This study can provide a reference for nature conservation and water resources management in Taiwan. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92119 |
DOI: | 10.6342/NTU202400522 |
全文授權: | 同意授權(限校園內公開) |
顯示於系所單位: | 土木工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-112-1.pdf 授權僅限NTU校內IP使用(校園外請利用VPN校外連線服務) | 5.52 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。