Muszynski, Grzegorz
(2021)
Pattern Recognition for Weather Phenomena in Climate Data.
PhD thesis, University of Liverpool.
|
Text
201248363_Feb2021.pdf - Unspecified Download (79MB) | Preview |
Abstract
Weather phenomena have long been objects of studies in atmospheric and climate science research. Studies on weather phenomena incorporate meteorological data, climate model simulations, and knowledge of physical processes of the Earth’s atmosphere. Understanding of the developing mechanisms, life cycles, and spatiotemporal dependencies of these phenomena requires accurately identifying them in space and time. Moreover, identifying weather phenomena in large-scale climate model simulations is critical for advancing our understanding of the Earth’s climate and risks of future extreme weather events. The main goal of this thesis is to design and develop pattern recognition methods that directly learn from examples of weather phenomena in climate data, rather than following heuristic algorithms containing threshold requirements on physical variables. In particular, we aim to classify and localise atmospheric river and blocking phenomena in global climate simulations and reanalysis data. In this thesis, we propose a novel pattern recognition method for identifying atmospheric river phenomena in climate datasets. This method consists of topological data analysis and machine learning methods. We demonstrate that the proposed method is reliable, robust, and achieves high accuracy. Also, we test the method on a wide range of spatial and temporal resolutions of global climate model outputs. We find that the method achieves the highest classification accuracy for low-resolution climate model outputs. Moreover, we propose a hierarchical pattern recognition method for identifying atmospheric blocking phenomena in climate reanalysis products. This pattern recognition method is based on deep convolutional neural networks. We demonstrate that the proposed method accurately detects and localises atmospheric blocks in climate reanalysis data. We also find that the method achieves higher accuracy for classification and lower estimation error for localisation of blocking phenomena in regions of the Northern Hemisphere than in regions of the Southern Hemisphere. Research outcomes presented in this thesis show that the proposed pattern recognition methods can be complementary tools to the existing identification methods of atmospheric rivers and blocks in climate data. In addition to that, the proposed methods offer automatic post-processing, quantitative assessment of climate datasets, and can facilitate analysis of the local impacts of weather phenomena on specific geographical areas.
| Item Type: | Thesis (PhD) |
|---|---|
| Divisions: | Faculty of Science and Engineering > School of Electrical Engineering, Electronics and Computer Science |
| Depositing User: | Symplectic Admin |
| Date Deposited: | 08 Mar 2021 15:35 |
| Last Modified: | 18 Jan 2023 22:58 |
| DOI: | 10.17638/03116349 |
| Supervisors: |
|
| URI: | https://livrepository.liverpool.ac.uk/id/eprint/3116349 |
Dimensions
Dimensions
