Gast ::
| Hauptseite > Publikationsdatenbank > Deep Leaning based data compression for climate data |
| Hauptseite > Publikationsdatenbank > Deep Leaning based data compression for climate data |
| Bachelor Thesis | FZJ-2025-02391 |
2024
This record in other databases:
Please use a persistent id in citations: doi:10.34734/FZJ-2025-02391
Abstract: With the amount of data generated in climate science from scientific models reachingmultiple terabytes per simulation, it becomes difficult to provide sufficient storagespace and achieve satisfactory data transfer rates. In order to be able to store moreinformation, data compression becomes increasingly relevant, as it allows the size ofthe data to be decreased significantly.In this work a novel approach to lossy data compression for climate data by usingthe deep learning model AtmoRep is being analysed. AtmoRep uses large-scalerepresentation learning to determine a general description of the highly complex,stochastic dynamics of the atmosphere which can be used to achieve a varietyof different tasks. Specifically, I want to test the accuracy of data reconstructionthat can be achieved with this model. This reconstructive ability can be used toefficiently compress data by purposefully removing parts of the data, which can bereconstructed using AtmoRep.To determine an optimal configuration of the AtmoRep model, this thesis proposesa series of different experiments where portions of temperature data are completelyremoved from the dataset at varying temporal positions and then reconstructedusing AtmoRep. The reconstructed data is subsequently analyzed and compared tothe original data to determine the quality of the data in relation to the compressionratio.For comparison, the results of AtmoRep are being compared to those of commonlyused lossless and lossy data compression algorithms for climate data.
|
The record appears in these collections: |
PDF