Driven by the increase of greenhouse gas emissions, climate change is causing significant shifts in the Earth's climatic patterns, profoundly affecting our oceans. In recent years, our capacity to monitor and understand the state and variability of the ocean has been significantly enhanced, thanks to improved observational capacity, new data-driven approaches, and advanced computational capabilities. Contemporary marine analyses typically integrate multiple data sources: numerical models, satellite data, autonomous instruments, and ship-based measurements. Temperature, salinity, and several other ocean essential variables, such as oxygen, chlorophyll, and nutrients, are among the most frequently monitored variables. Each of these sources and variables, while providing valuable insights, has distinct limitations in terms of uncertainty, spatial and temporal coverage, and resolution. The application of deep learning offers a promising avenue for addressing challenges in data prediction, notably in data reconstruction and interpolation, thus enhancing our ability to monitor and understand the ocean. This thesis proposes and evaluates the performances of a variety of neural network architectures, examining the intricate relationship between methods, ocean data sources, and challenges. A special focus is given to the biogeochemistry of the Mediterranean Sea. A primary objective is predicting low-sampled biogeochemical variables from high-sampled ones. For this purpose, two distinct deep learning models have been developed, each specifically tailored to the dataset used for training. Addressing this challenge not only boosts our capability to predict biogeochemical variables in the highly heterogeneous Mediterranean Sea region but also allows the increase in the usefulness of observational systems such as the BGC-Argo floats. Additionally, a method is introduced to integrate BGC-Argo float observations with outputs from an existing deterministic marine ecosystem model, refining our ability to interpolate and reconstruct biogeochemical variables in the Mediterranean Sea. As the development of novel neural network methods progresses rapidly, the task of establishing benchmarks for data-driven ocean modeling is far from complete. This work offers insights into various applications, highlighting their strengths and limitations, besides highlighting the importance relationship between methods and datasets.

Driven by the increase of greenhouse gas emissions, climate change is causing significant shifts in the Earth's climatic patterns, profoundly affecting our oceans. In recent years, our capacity to monitor and understand the state and variability of the ocean has been significantly enhanced, thanks to improved observational capacity, new data-driven approaches, and advanced computational capabilities. Contemporary marine analyses typically integrate multiple data sources: numerical models, satellite data, autonomous instruments, and ship-based measurements. Temperature, salinity, and several other ocean essential variables, such as oxygen, chlorophyll, and nutrients, are among the most frequently monitored variables. Each of these sources and variables, while providing valuable insights, has distinct limitations in terms of uncertainty, spatial and temporal coverage, and resolution. The application of deep learning offers a promising avenue for addressing challenges in data prediction, notably in data reconstruction and interpolation, thus enhancing our ability to monitor and understand the ocean. This thesis proposes and evaluates the performances of a variety of neural network architectures, examining the intricate relationship between methods, ocean data sources, and challenges. A special focus is given to the biogeochemistry of the Mediterranean Sea. A primary objective is predicting low-sampled biogeochemical variables from high-sampled ones. For this purpose, two distinct deep learning models have been developed, each specifically tailored to the dataset used for training. Addressing this challenge not only boosts our capability to predict biogeochemical variables in the highly heterogeneous Mediterranean Sea region but also allows the increase in the usefulness of observational systems such as the BGC-Argo floats. Additionally, a method is introduced to integrate BGC-Argo float observations with outputs from an existing deterministic marine ecosystem model, refining our ability to interpolate and reconstruct biogeochemical variables in the Mediterranean Sea. As the development of novel neural network methods progresses rapidly, the task of establishing benchmarks for data-driven ocean modeling is far from complete. This work offers insights into various applications, highlighting their strengths and limitations, besides highlighting the importance relationship between methods and datasets.

MACHINE LEARNING APPLICATIONS TO DATA RECONSTRUCTION IN MARINE BIOGEOCHEMISTRY / Pietropolli, Gloria. - (2024 Mar 21).

MACHINE LEARNING APPLICATIONS TO DATA RECONSTRUCTION IN MARINE BIOGEOCHEMISTRY.

PIETROPOLLI, GLORIA
2024-03-21

Abstract

Driven by the increase of greenhouse gas emissions, climate change is causing significant shifts in the Earth's climatic patterns, profoundly affecting our oceans. In recent years, our capacity to monitor and understand the state and variability of the ocean has been significantly enhanced, thanks to improved observational capacity, new data-driven approaches, and advanced computational capabilities. Contemporary marine analyses typically integrate multiple data sources: numerical models, satellite data, autonomous instruments, and ship-based measurements. Temperature, salinity, and several other ocean essential variables, such as oxygen, chlorophyll, and nutrients, are among the most frequently monitored variables. Each of these sources and variables, while providing valuable insights, has distinct limitations in terms of uncertainty, spatial and temporal coverage, and resolution. The application of deep learning offers a promising avenue for addressing challenges in data prediction, notably in data reconstruction and interpolation, thus enhancing our ability to monitor and understand the ocean. This thesis proposes and evaluates the performances of a variety of neural network architectures, examining the intricate relationship between methods, ocean data sources, and challenges. A special focus is given to the biogeochemistry of the Mediterranean Sea. A primary objective is predicting low-sampled biogeochemical variables from high-sampled ones. For this purpose, two distinct deep learning models have been developed, each specifically tailored to the dataset used for training. Addressing this challenge not only boosts our capability to predict biogeochemical variables in the highly heterogeneous Mediterranean Sea region but also allows the increase in the usefulness of observational systems such as the BGC-Argo floats. Additionally, a method is introduced to integrate BGC-Argo float observations with outputs from an existing deterministic marine ecosystem model, refining our ability to interpolate and reconstruct biogeochemical variables in the Mediterranean Sea. As the development of novel neural network methods progresses rapidly, the task of establishing benchmarks for data-driven ocean modeling is far from complete. This work offers insights into various applications, highlighting their strengths and limitations, besides highlighting the importance relationship between methods and datasets.
21-mar-2024
COSSARINI, GIANPIERO
Manzoni, Luca
36
2022/2023
Settore INF/01 - Informatica
Università degli Studi di Trieste
File in questo prodotto:
File Dimensione Formato  
PhD_Thesis_Pietropolli_def.pdf

accesso aperto

Descrizione: PhD thesis
Tipologia: Tesi di dottorato
Dimensione 15.97 MB
Formato Adobe PDF
15.97 MB Adobe PDF Visualizza/Apri
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11368/3071880
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact