2023-04-012024-05-13https://scholars.lib.ntu.edu.tw/handle/123456789/653849地球科學領域中的古氣候學研究用以幫助我們鑑古知今並預測未來的氣候變化，進而提出人類活動相應的措施。透過地球物理、地球化學、古生物學的分析，我們可以觀察在地質紀錄中所透露的古氣候資訊，之中常見的有記錄沈積環境變化的沈積相以及表現碳循環紀錄的碳酸鹽、有機碳、無機碳濃度變化，若能搭配年代的支持便能拼湊出過去萬年來地球氣候變化的歷史。常規的沈積相辨識需仰賴沈積學家對沈積物的構造、顏色、粒徑等的觀察並輔以沉積學知識與經驗做出判斷，雖然這個方法相對快速且符合沈積學原理；但具有受觀察者影響的差異且屬於定性分析，不易量化作為後續與其他研究之比較或進一步分析。雖然能夠透過多位沈積學家共同辨識以降低觀察者的偏差，但如此一來所增加的時間成本便失去此方法快速之優點。另外也可以透過實驗手法妥善處理沉積物，在予以測量其中化學成分與濃度的變化以達到量化的目的；但這些常規定量方法亦將大幅地增加人力與時間成本，使得研究不符合效益或測量的解析度須作出妥協退讓。拜科技進步所賜，新興的岩芯掃描技術如X光螢光岩芯掃描儀以及光學影像，提供了可以克服上述觀察者偏差與成本限制的問題，這些技術能夠非破壞性地量測出沉積物高解析度（半）定量的物理、化學成分變化，其解析度可以達到微米的尺度，而且測量所需之時間可以從常規定量方法的數天縮短至數小時。惟因其省去樣本前處理，測量值需要更為複雜的校正以達到常規定量技術的量化水平。近年來機器學習隨著電腦計算能力的提升以及數據獲得的難度降低而廣泛受到各學門的青睞。機器學習是屬電腦科學與統計數學的結合技術，透過演算法學習大量的數據與結果建立模型，並依此模型來預測新數據的結果為何，如此便可免去手動建立複雜判斷條件的模型。眾多學門如生物學及腦神經科學等已陸續開始引入機器學習作為研究的輔助技術，於地球科學領域雖有相關應用但仍屬初期試探階段，仍有許多科學問題值得投入機器學習的跨領域合作進而獲得新突破。本研究計畫視機器學習作為橋樑結合常規研究方法與岩芯掃描技術的潛力，兩種方法的優點得以被保留同時減少既存缺點。來自德國瓦登海岸以及西北太平洋深海的大量珍貴岩芯皆擁有常規研究及岩芯掃描數據，我們期許利用機器學習從這些數據中建立模型來達到兩個突破，一為使用岩芯掃描之X光螢光頻譜、化學元素強度、光學影像進行沈積相的自動辨識，二為校正岩芯掃描之X光螢光頻譜產生沉積物化學成分濃度值以大幅提升量化之解析。 Geoscience provides us with paths to explore the earth. Paleoclimate is one of the paths that help us to predict the natural influence on human activities in the near future. Many measurements, such as geophysical, geochemical, and faunistic analyses, have been developed to observe the information recorded in geological archives. Sediment facies, defined as a series of specific deposition conditions, indicate paleoenvironmental variation. Bulk chemistry in sediments, including carbonate, total carbon (TC), and total organic carbon (TOC), depicts chemical and biological processes during sedimentation. If together with chronology, the measurements playessential rolesin constructing detailed paleoclimate. Classic discrimination of sediment facies relies on sedimentologists` judgment, combining their macroscopically or microscopically observations on sediment with their sedimentary knowledge and experience. Although this way provides a quick and logical decision based on sedimentary, it has some drawbacks like observer-dependent, qualitative property and not easy to re-evaluate. It can be carried out by multiple researchers independently to minimized observer-bias, which yet makes it labor-intensive. The conventional approach of measuring bulk chemistry provides a quantitative perspective that solves the observer-dependence. However, it often requires intensive labor and time. The measuring resolution is also limited. Novel down-core scanning techniques (X-ray fluorescence core scanner (XRF-CS) and photography) provide a work-around by rapidly producing high-resolution and (semi)quantitative measurements. The scanning time can be reduced from days to just hours. The measuring resolution can reach the micron scale in Micro-XRF-CS. Qualitative assessment can be replaced by (semi)quantitative, which reduces the effect of observer-dependence and is stored digitally. Nevertheless, this (semi)quantitative property sometimes causes unneglectable uncertainty in studies, so the measurements often require data analysis and calibration. As an up-and-coming discipline sitting at the intersection of computer science and statistics, machine learning (ML) has proven its capability to assist research questions from extensive disciplines. It intends to train computers using exemplary data, often a large dataset, instead of manually craft complex decision rules to solve a given problem. Some studies have introduced ML into Geosciences, but still in a relatively trial scale comparing to other disciplines, such as biology and neuroscience. This proposal sees ML`s potential as a bridge to connect the conventional methods and the scanning techniques in Geosciences. We expect to complement the advantages of both approaches and lessen their disadvantages by applying sufficient materials from the coast of the Wadden Sea and the deep sea of the Northwest Pacific Ocean. These materials, sharing valuable geological interests, have both the conventional and scanning measurements. As a result, the proposal addresses two goals to be achieved: building models by ML techniques to (1) discriminate sediment facies using XRF spectra, elemental intensities, or photographs and (2) enhance measuring resolution of bulk chemistry by calibrating XRF spectra. After the bridge is successfully established, the contributions shall mature the ML approach in Geosciences and give researchers more strength to further explore the earth`s interests. Moreover, this proposal`s cooperative nature is anticipated to broaden the network between Taiwanese and German scientific communities. The proposal is planned in three years schedule.機器學習; X光螢光頻譜; 掃描技術;machine learning; XRF spectra; scanning techniques