國立虎尾科技大學 |

Democratizing Access to Extensive Climate Datasets via Deep Learning-Powered Techniques.

紀錄類型:	書目-語言資料,印刷品 : Monograph/item
正題名/作者:	Democratizing Access to Extensive Climate Datasets via Deep Learning-Powered Techniques./
作者:	Kurihana, Takuya.
出版者:	Ann Arbor : ProQuest Dissertations & Theses, : 2024,
面頁冊數:	167 p.
附註:	Source: Dissertations Abstracts International, Volume: 85-09, Section: B.
Contained By:	Dissertations Abstracts International85-09B.
標題:	Computer engineering. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30990391
ISBN:	9798381943535

Democratizing Access to Extensive Climate Datasets via Deep Learning-Powered Techniques.
Kurihana, Takuya.

Democratizing Access to Extensive Climate Datasets via Deep Learning-Powered Techniques. - Ann Arbor : ProQuest Dissertations & Theses, 2024 - 167 p.

Source: Dissertations Abstracts International, Volume: 85-09, Section: B.

Thesis (Ph.D.)--The University of Chicago, 2024.

This item must not be sold to any third party vendors.

Artificial Intelligence (AI) for science (AI4Science) develops cutting-edge AI algorithms and High-Performance Computing (HPC) to advance the frontier of science through the discovery of new scientific knowledge. However, hundreds of terabytes to petabytes of vast volumes of climate science datasets, including multispectral satellite instruments and numerical simulations, pose a challenge to processing and extracting useful information: Satellite instruments have captured cloud structure, size distributions, and radiative properties at a near-daily cadence over the past decades. High-resolution numerical climate and weather simulations have contributed to understanding the complicated interactions and feedback of Earth systems. These observations and simulations should help understand cloud and climate responses, but the complexity and size of this dataset have left it under-utilized. To aid the challenge and achieve the democratization of large volumes of climate science datasets by lowering a barrier to access to the core data, neural network-based approaches using self-supervised and supervised learning are promising solutions. In the first part of this thesis, I introduce rotationally invariant cloud clustering (RICC) that combines rotationally invariant autoencoder and hierarchical agglomerative clustering to generate unique new AI-generated cloud classes. Clusters produced from RICC detect meaningful distinctions between cloud textures, using only raw multispectral imagery as an input without reliance on location, time, derived physical properties, or pre-designated class definitions. Having RICC, I create a unique new cloud dataset, the AI-driven cloud classification atlas (AICCA), which clusters 23 years of ocean images from the Moderate Resolution Imaging Spectroradiometer (MODIS) on NASA’s Aqua and Terra instruments –190 million of roughly 100 km x 100 km patches (128 x 128 pixels) - into 42 AI-generated cloud class labels. AICCA translates 872 TB of satellite images into 56 GB of class labels, metadata, and 13 cloud physical parameters. I finally demonstrate that RICC and AICCA apply to build a universal workflow to evaluate the bias exhibited by simulated clouds from high-resolution cloud models. In the second part of the thesis, I present various neural network-based methodologies designed to deliver compressed or upscaled resolution of climate datasets, catering to the varied requirements of climate analysis. Clustering autoencoder that incorporates an online clustering algorithm offers a data-driven climate classification approach without subjective definitions. This method addresses a computational limitation in off-line clustering in RICC and compresses 70 years of GFDL-ESM2G climate simulation at 0.125 spatial resolution over the Continental United States under multiple warming scenarios, reducing it to a lower-dimensional space by a factor of 660,000. Additionally, I develop a physics-informed generative adversarial network utilizing self-attention computation to capture three-dimensional weather dynamics. The network super-resolves the three-dimensional wind data by upscaling the resolution by a factor of nine, ultimately aiming to provide accurate tracing and monitoring of greenhouse gas emissions. The super-resolution generative model compensates for the absence of high-resolution simulation outputs and saves computational time.

ISBN: 9798381943535Subjects--Topical Terms:

569006
Computer engineering.
Subjects--Index Terms:

AI4Science

Democratizing Access to Extensive Climate Datasets via Deep Learning-Powered Techniques.
LDR:04578nam a2200373 4500 001 1121938
005 20240624103718.5
006 m o d
007 cr#unu||||||||
008 240823s2024 ||||||||||||||||| ||eng d
020 $a 9798381943535
035 $a (MiAaPQ)AAI30990391
035 $a AAI30990391
040 $a MiAaPQ $c MiAaPQ
100 1 $a Kurihana, Takuya. $3 1437845
245 1 0 $a Democratizing Access to Extensive Climate Datasets via Deep Learning-Powered Techniques.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2024
300 $a 167 p.
500 $a Source: Dissertations Abstracts International, Volume: 85-09, Section: B.
500 $a Advisor: Foster, Ian.
502 $a Thesis (Ph.D.)--The University of Chicago, 2024.
506 $a This item must not be sold to any third party vendors.
520 $a Artificial Intelligence (AI) for science (AI4Science) develops cutting-edge AI algorithms and High-Performance Computing (HPC) to advance the frontier of science through the discovery of new scientific knowledge. However, hundreds of terabytes to petabytes of vast volumes of climate science datasets, including multispectral satellite instruments and numerical simulations, pose a challenge to processing and extracting useful information: Satellite instruments have captured cloud structure, size distributions, and radiative properties at a near-daily cadence over the past decades. High-resolution numerical climate and weather simulations have contributed to understanding the complicated interactions and feedback of Earth systems. These observations and simulations should help understand cloud and climate responses, but the complexity and size of this dataset have left it under-utilized. To aid the challenge and achieve the democratization of large volumes of climate science datasets by lowering a barrier to access to the core data, neural network-based approaches using self-supervised and supervised learning are promising solutions. In the first part of this thesis, I introduce rotationally invariant cloud clustering (RICC) that combines rotationally invariant autoencoder and hierarchical agglomerative clustering to generate unique new AI-generated cloud classes. Clusters produced from RICC detect meaningful distinctions between cloud textures, using only raw multispectral imagery as an input without reliance on location, time, derived physical properties, or pre-designated class definitions. Having RICC, I create a unique new cloud dataset, the AI-driven cloud classification atlas (AICCA), which clusters 23 years of ocean images from the Moderate Resolution Imaging Spectroradiometer (MODIS) on NASA’s Aqua and Terra instruments –190 million of roughly 100 km x 100 km patches (128 x 128 pixels) - into 42 AI-generated cloud class labels. AICCA translates 872 TB of satellite images into 56 GB of class labels, metadata, and 13 cloud physical parameters. I finally demonstrate that RICC and AICCA apply to build a universal workflow to evaluate the bias exhibited by simulated clouds from high-resolution cloud models. In the second part of the thesis, I present various neural network-based methodologies designed to deliver compressed or upscaled resolution of climate datasets, catering to the varied requirements of climate analysis. Clustering autoencoder that incorporates an online clustering algorithm offers a data-driven climate classification approach without subjective definitions. This method addresses a computational limitation in off-line clustering in RICC and compresses 70 years of GFDL-ESM2G climate simulation at 0.125 spatial resolution over the Continental United States under multiple warming scenarios, reducing it to a lower-dimensional space by a factor of 660,000. Additionally, I develop a physics-informed generative adversarial network utilizing self-attention computation to capture three-dimensional weather dynamics. The network super-resolves the three-dimensional wind data by upscaling the resolution by a factor of nine, ultimately aiming to provide accurate tracing and monitoring of greenhouse gas emissions. The super-resolution generative model compensates for the absence of high-resolution simulation outputs and saves computational time.
590 $a School code: 0330.
650 4 $a Computer engineering. $3 569006
650 4 $a Computer science. $3 573171
653 $a AI4Science
653 $a Cloud classification
653 $a Deep learning
690 $a 0984
690 $a 0464
690 $a 0800
710 2 $a The University of Chicago. $b Computer Science. $3 845504
773 0 $t Dissertations Abstracts International $g 85-09B.
790 $a 0330
791 $a Ph.D.
792 $a 2024
793 $a English
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30990391