Application of data- driven models in geophysics

[F. Arnaut]

129 стр.

Geo-nauke - Geofizika / Geosciences - Geophysics

Modeli vođeni podacima karakterišu se ubrzanim razvojem zbog široke pristupačnosti i unapređenja sposobnosti procesora, softverskih paketa i računarske memorije. Primena metoda mašinskog učenja, prognoziranja i analize vremenskih serija aktivno napreduje unutar šire naučne oblasti nauke o podacima. Potencijalna područja primene ovih metoda na geofizičke, geološke i podatke koji se odnose na blisku Zemljinu okolinu su brojna i raznovrsna. Doktorska disertacija prikazuje kombinovane rezultate različitih primena prethodno pomenutih metoda, i to: primena metoda prognoziranja vremenskih serija na podatke koncentracije zagađujućih materija u vazduhu (Fejsbuk Profet model), primena metoda mašinskog učenja za imputaciju podataka koncentracije zagađujućih materija u vazduhu (dvosmerna imputacija podataka), primena metoda klasifikacije mašinskog učenja za prostornu klasifikaciju ofiolita istočne Vardarske zone (Severna Makedonija), primena metoda mašinskog učenja na podatke signala vrlo niskih frekvencija koji se prostiru subjonosferski za detekciju poremećaja amplitude pomenutih signala, primena metoda mašinskog učenja za dobijanje talasovodnih parametara oblasti D Zemljine jonosfere tokom poremećenih jonosferskih uslova usled solarnih flerova i primena statističkih metoda na podatke magnetne susceptibilnosti dobijene od uzoraka materijala sa flotacijskog jalovišta rudnika „Rudnik“ (Srbija). Svaki od prethodnih primera prikazuje primenu različitih metoda nauke o podacima (mašinsko učenje i metode prognoziranja vremenskih serija) i primenjene statistike na realne podatke, kako bi se dobile informacije koje nije moguće dobiti konvencionalnim metodama. Disertacija takođe prikazuje značaj budućih unapređenja i primene ovih metoda na geo-podatke, koji mogu imati veliki značaj za istraživače i industriju u datim domenima.

Data-driven models have rapidly expanded due to widespread accessibility and advancements in processing capabilities, software packages, and computer memory. The utilization of machine learning and time-series forecasting and analysis techniques is actively progressing within the broader scientific domain of data science. The potential application of these methods to geophysical, geological, and near-Earth physics datasets is extensive and diverse. This PhD dissertation presents the combined results from various applications, specifically: the utilization of time-series forecasting techniques (Facebook’s Prophet algorithm) on particulate matter concentration data; the employment of machine learning methods to impute missing observations (bi-directional data imputation) for particulate matter concentration data; the application of machine learning classification techniques for the spatial classification of ophiolites in the East Vardar Ophiolite Zone, North Macedonia; the implementation of machine learning methods on subionospheric very low frequency signal data for amplitude anomaly detection; the use of machine learning methods to derive D region ionospheric waveguide parameters under disturbed ionospheric conditions due to solar flares; and the application of statistical methods on magnetic susceptibility data acquired from a flotation mine tailing material from the mine “Rudnik” (Serbia). Each of the aforementioned examples illustrates the application of diverse data science (machine learning and time- series forecasting methods) and applied statistical methodologies on real-world data to extract insights that are not conventionally attainable. The dissertation emphasizes the significance of future advancements and applications of these methods for geo-data, as they are highly beneficial for researchers and industry professionals in these domains., Data-driven models have rapidly expanded due to widespread accessibility and advancements in processing capabilities, software packages, and computer memory. The utilization of machine learning and time-series forecasting and analysis techniques is actively progressing within the broader scientific domain of data science. The potential application of these methods to geophysical, geological, and near-Earth physics datasets is extensive and diverse. This PhD dissertation presents the combined results from various applications, specifically: the utilization of time-series forecasting techniques (Facebook’s Prophet algorithm) on particulate matter concentration data; the employment of machine learning methods to impute missing observations (bi-directional data imputation) for particulate matter concentration data; the application of machine learning classification techniques for the spatial classification of ophiolites in the East Vardar Ophiolite Zone, North Macedonia; the implementation of machine learning methods on subionospheric very low frequency signal data for amplitude anomaly detection; the use of machine learning methods to derive D region ionospheric waveguide parameters under disturbed ionospheric conditions due to solar flares; and the application of statistical methods on magnetic susceptibility data acquired from a flotation mine tailing material from the mine “Rudnik” (Serbia). Each of the aforementioned examples illustrates the application of diverse data science (machine learning and time- series forecasting methods) and applied statistical methodologies on real-world data to extract insights that are not conventionally attainable. The dissertation emphasizes the significance of future advancements and applications of these methods for geo-data, as they are highly beneficial for researchers and industry professionals in these domains.

mašinsko učenje, prognoziranje vremenskih serija, prostorno prognoziranje, nauka o podacima, analiza (geo)podataka.

machine learning, time- series forecasting, spatial forecasting, data science, (geo)data analysis.

550.3:004.85(043.85)

Tekst