| 摘要: | 地電阻和透地雷達（GPR）方法運用已超過一世紀。事實上，最早的地電阻法研究約在1920年代，重點探討一維（1D）模型中不同深度的電導率。另一方面，GPR最初是在1930年代作為評估冰川厚度的工具。近年來，這些方法已經顯著發展，有更強大且高效的設備和數據收集儀器的出現。隨著有更多技術成熟的儀器，更需要關注如何有效地管理、分析和解釋這些方法收集的數據。與數據分析相關的問題雖已存在很久，卻常常被研究人員忽視。首先，大多數反演軟件的價格高昂，使許多研究人員無法負擔進行研究。其次，在分析和解釋地電測量獲得的主要和次要數據時，仍然存在挑戰，特別是在處理時間序列數據、進行正演建模和將獲得的數據轉化為其他訊息或參數（如岩性或其他物理參數）時。第三，該領域缺乏有效整合多種方法的電阻數據的成熟技術，許多研究人員仍然依賴於過時的方法，如將來自各種測量技術的結果並排比較，或疊加彼此的圖像以找到它們之間的聯繫。第四，三維（3D）建模的精度存在問題，雖然研究人員通常使用商業軟件處理3D模型，但這些工具可能會引入過度擬合和數據的過度插值等問題。此外，這些軟件有時難以充分處理模型的邊界條件，導致缺乏彈性，甚至過度解示。第五，數據處理及解釋耗時且需要大量人力進行，特別是在處理大量數據集並嘗試同時識別多個不同特徵時，例如在雷達圖中判視鋼筋位置。為此需要研究更先進的方法來有效解決這個問題。最後，數據分析的定量方法存在顯著不足。許多以前的研究主要依賴於對來自二維（2D）數據的結果的視覺解釋，如標示結構、異常或目標，而沒有採用定量方法。這是一個重大疏漏，因為作為地球物理學家，擁有數值及其相應的物理意義在進行研究時至關重要。
本研究旨在透過機器學習（ML）並結合Python的模組和視覺化，引入精密的處理步驟和數據分析，解決判識上的挑戰。首先，我們提出了一種以Python為基礎的方法利用機器學習來克服高成本的反演軟體，本研究使用監督式機器學習（SML）技術，如隨機森林、支持向量機和線性回歸，將結果與傳統的最小二乘法進行比較，隨機森林算法在正演模擬的數據和現地數據的均方根誤差較低，R2值較高，顯示隨機森林算法有較好的運算潛力。其次，為了加強對於電阻率數據的解釋，本研究應用了非監督式機器學習（UML）方法，其中包括聚合式階層分群法（HAC）和K平均演算法，用於主要和次要的電阻率資料。在主數據集中，解釋用現地和正演數據，經反演出的電阻率剖面，而在次要數據集中，我們將電阻率數據轉化為地下水參數，揭示了測量區域內獨特的模式。這種方法使我們對地下水隨季節變化模式有了新的理解。本研究成功地利用統計方法中的數據同化來整合多種地電數據，超越了傳統的直接比較方法，因為該方法會忽略每次測量的電阻率範圍。此外，本研究透過Python的模組和視覺化，逐步開發了一個避免了過度擬合和過度插值的3D電阻率模型。這與商業軟件形成了對比，商業軟體模型的邊界和地形是預設的。我們更進一步利用SML，將3D電阻率模型轉換為3D似地質模型，其中的鑽井資料作為模型的真實數據。其次，應用於電阻率成像（ERI）的ML程序被擴展到透地雷達（GPR）數據分析。實現了自動化解釋雷達圖，包括檢測襯砌、空洞和鋼筋。本研究也建立了GPR的3D模型，解決相位差異，改善視覺化的結果。更進一步，我們以定量方式加強了ERI和GPR的數據分析。對於ERI，應用了非監督式機器學習來描繪2D模型的結構邊界，並透過2D正演模擬驗證了該方法。對於GPR方面，該方法成功應用在對於地下水的定量研究，利用各種統計模型估算土壤含水量和地下水位。
總體而言，本研究針對多年對於地電阻法和透地雷達法應用中所遇到的挑戰提供了許多有效的解決方案。對於現地數據進行進階分析也已用正演模擬得到驗證，為解決地電阻法和透地雷達法研究中過去的問題提供了見解，期待此研究成果能有助於引領未來的研究工作。
;Geoelectrical and Ground Penetrating Radar (GPR) methods have been in use for over a century. In fact, the first attempt at geoelectrical research took place in the 1920s, focusing on electrical conductivity at different depths in a one-dimensional (1D) model. On the other hand, GPR was initially introduced as a tool for evaluating glacier thickness in the 1930s. Over the years, these methods have evolved significantly, resulting in more powerful and efficient equipment and data acquisition instruments. With well-developed instruments now available, the primary challenge that demands more attention is how to effectively manage, analyze, and interpret the acquired data from these methods. Several issues related to data analysis have persisted for an extended period, often overlooked by researchers.
Firstly, the majority of inversion software comes at a steep cost, rendering it unaffordable for many researchers to conduct their studies. Secondly, challenges persist in analyzing and interpreting primary and secondary data derived from geoelectrical measurements, particularly when dealing with time series data, conducting forward modeling, and transforming acquired data into meaningful information, such as lithology or other physical parameters. Thirdly, the field lacks a well-developed technique for effectively integrating multi-geoelectrical data. Many researchers still rely on outdated methods, such as comparing results from various measurement techniques side by side or overlaying each other′s images to find links between them. Fourthly, there are issues concerning the accuracy of three-dimensional (3D) modeling. While researchers often turn to commercial software for handling 3D models, these tools can introduce problems like overfitting and excessive interpolation of the data. Furthermore, such software sometimes struggles to adequately address the model′s boundary conditions, resulting in inflexibility and misinterpretation. Fifthly, there is a significant demand for time and labor-intensive data processing and interpretation, especially when dealing with numerous datasets and attempting to identify multiple distinct features simultaneously, such as locating reinforcement bars or rebar in radargrams. Advanced methods need to be explored to effectively address this issue. Finally, there is a significant deficiency in quantitative methods for data analysis. Many previous studies primarily rely on visual interpretations of results obtained from two-dimensional (2D) data, such as delineating structures, anomalies, or targets, without incorporating quantitative approaches. This oversight is significant since, as geophysicists, having numerical values along with their corresponding physical meanings is of utmost importance in conducting research.
To address the identified challenges, this study aims to introduce sophisticated processing steps and data analysis through the integration of Machine Learning (ML) with Python-based modeling and visualization.
Firstly, a Python-based approach coupled with machine learning is proposed to tackle the high-cost inversion software. Supervised Machine Learning (SML) techniques, such as Random Forest, Support Vector Machine, and Linear Regression, are employed. The results are compared with the conventional Least Square Method, revealing that the Random Forest algorithm exhibits promise with lower RMSE and higher R2 for both forward modeling simulation data and field data. Secondly, to enhance the interpretation of resistivity data, Unsupervised Machine Learning (UML) methods, including Hierarchical Agglomerative Clustering (HAC) and K-means algorithms, are applied to both primary and secondary resistivity data. In the primary dataset, the inverted resistivity section for both field and forward modeling data is interpreted, while in the secondary dataset, resistivity data transformed into groundwater parameters reveal unique patterns in the survey area. This approach leads to a novel understanding of groundwater behavior within seasonal fluctuations. Thirdly, the study successfully integrates multi-geoelectrical data using data assimilation in statistical methods, moving beyond the traditional direct comparison approach that neglects the resistivity range of each measurement. Fourthly, by employing Python-based modeling and visualization with a detailed step-by-step procedure, a 3D resistivity model is developed without overfitting or excessive interpolation. This is in contrast to commercial software, where the boundary and topography of the model are predefined. Moreover, utilizing SML, the 3D resistivity model is converted into a 3D apparent geological model, where borehole data act as ground truth data. Fifthly, the ML procedures developed for Electrical Resistivity Imaging (ERI) are extended to Ground Penetrating Radar (GPR) data analysis. Automation of radargram interpretation, including the detection of lining, voids, and rebar, is achieved. A 3D model for GPR data is constructed, addressing phase differences for improved visualization. Finally, data analysis is enhanced quantitatively for both ERI and GPR data. For ERI data, Unsupervised Machine Learning is applied to delineate structure boundaries in 2D data, confirming the success of this method through 2D forward modeling simulation. For GPR data, this method is successfully implemented for quantitative groundwater studies, estimating soil water content and groundwater table using various statistical models.
In summary, this dissertation provides valuable solutions to the challenges encountered in employing geoelectrical and GPR methods for over the years. The advanced procedures applied to field data are validated through forward modeling simulation, offering insights that could address past issues in geoelectrical and GPR studies. Ultimately, these findings contribute to navigating future research endeavors in the field. |
