利用機器學習探討沉積學與古海洋學研究框架下之自動化沉積相辨識與高解析地球化學量化應用

Abstract

解析地球系統的運作機制能夠促進我們在科學以及人類社會上的進步與永續，其中一項重要研究材料為記錄著短中長期氣候、地形、生物與人類活動訊號的沉積物紀錄。結合地球科學與電腦科學兩大領域的進展，本論文展示三項解決常見於地球系統中沉積學與古海洋學研究之問題的跨領域研究。第一項研究彙整多種採自德國瓦登海域之岩心掃描數據進行數據分析以解析不同沉積相之特性，結果顯示這些沉積物紀錄中的掃描數據確實於不同沉積相中具有不同之特徵，故從中我們得知沉積學家的觀察能夠再現於掃描數據。於是第二項研究「沉積相自動辨識」接續進行，所挑選之岩心及沉積相數量較第一項研究顯著提升，目標在於透過機器學習技術建立能藉由讀取X光螢光岩心掃描儀（以下簡稱XRF）所產生之元素資料進行沉積物之沉積相辨識的模型。研究流程將一系列之特徵轉換與機器學習演算法進行組合測試並模擬出類似沉積學家觀察模式之資料轉換，透過融入沉積學原理特製化之模型能力檢驗，獲得一組最佳化具78%準確度之模型。除了具沉積相辨識能力，該模型亦可標記出模型較不具信心之沉積物區段以便予沉積學家投入其專業進行判斷，如此便可提升辨識量能且不影響整體品質。第三項研究則著重於產生高解析度之沉積物定量化學性質（碳酸鈣及總有機碳），本方法透過機器學習演算法直接對XRF所產出之原始光譜數據以及定量數據進行非線性迴歸建模，從此便可使用該模型對快速產生之高解析XRF數據進行量化求得碳酸鈣及總有機碳之分佈；此外，本研究納入太平洋高緯度多處之沉積物紀錄以突破過去常受限於單一井位或地區之應用性，其碳酸鈣及總有機碳量化準確度（R2）分別為0.96與0.78。總結來說，機器學習技術能夠提升資料分析以及自動化之能力，而岩心掃描技術則提供快速且高解析度之多項測量，本論文提供結合這兩種技術之方法藍圖以期能夠突破過去沉積物研究常見之費工耗時與主觀影響，進而獲得更全面之沉積物紀錄資訊。

Studying the mother Earth to understand and predict its system process facilitates the evolution of science and human society toward a progressive and sustainable stage. One of the key research materials is sediment archive, which records various short-and long-term historical information, such as climatic, biological, geomorphological and human activity variations. With the progress in geoscience and computer science, this thesis presents three interdisciplinary approaches to solve practical problems in sedimentological and paleoceanographical investigations. The first study compiles different core scanning data (magnetic susceptibility, X-ray computed tomography, elemental profiles, digital photography) to characterize a priori classified sediment core sections recovered from the Wadden Sea region. The results confirm that the description of human-recognized sediment facies can be reproduced using the scanning data, which gives a promising hint to a further step: automatic sediment facies classification. The second study increases the data scale by covering more core sections and sediment facies to develop a machine learning (ML) model that classifies sediments into facies by reading elemental profiles acquired from X-ray fluorescence (XRF) core scanning. A series of feature engineering and ML algorithms are tested to find the optimal solution. As a result, a simple but powerful model is proposed to simulate sedimentologists’ observational behavior and have promising performance (78% accuracy), which is supported by a tailor-made evaluation involving sedimentary knowledge. Furthermore, the model can highlight critical sections of sediments requiring sedimentologists’ expertise. This provides an increased capability of classification without losing accuracy. The third study focuses on obtaining cost-efficient high-resolution bulk chemistry measurement (CaCO3 and total organic carbon) by quantifying XRF spectra using ML. This novel approach of using XRF spectra directly and enhanced regression power of ML eliminates manual bias and increases input information. Meanwhile, the previous limitation of data coverage is lifted by including multi-regional data (high latitude sectors of Pacific Ocean). The outcome is carefully evaluated using cross-validation, test set, and case study, with R2 of CaCO3: 0.96 and TOC: 0.78 from the test set. In conclusion, ML increases the capability of data analysis and automation. Sediment core scanning provides thorough but rapid measurement in high spatial resolution. This thesis offers generalizable methodological blueprints for integrating these two techniques to release the wealth of sedimentary information from the shackles of expensive and observer-dependent data in the past.