Imperial College London
Portrait Picture

Dr Julia E. Stawarz

Faculty of Natural Sciences, Department of Physics

Academic Visitor
 
 
 
//

Contact

 

+44 (0)20 7594 7766j.stawarz

 
 
//

Location

 

6M71Huxley BuildingSouth Kensington Campus

//

Summary

 

Publications

Citation

BibTex format

@article{Smith:2024:10.1029/2023JA032096,
author = {Smith, AW and Rae, IJ and Stawarz, JE and Sun, WJ and Bentley, S and Koul, A},
doi = {10.1029/2023JA032096},
journal = {Journal of Geophysical Research: Space Physics},
title = {Automatic Encoding of Unlabeled Two Dimensional Data Enabling Similarity Searches: Electron Diffusion Regions and Auroral Arcs},
url = {http://dx.doi.org/10.1029/2023JA032096},
volume = {129},
year = {2024}
}
Download

RIS format (EndNote, RefMan)

TY  - JOUR
AB  - Critically important phenomena in Earth’s magnetosphere often occur briefly, or in small spatial regions. These processes are sampled with orbiting spacecraft or by fixed ground observatories and so rarely appear in data. Identifying such intervals can be an incredibly time consuming task. We apply a novel, powerful method by which two dimensional data can be automatically processed and embeddings created that contain key features of the data. The distance between embedding vectors serves as a measure of similarity. We apply the state-of-the-art method to two example datasets: MMS electron velocity distributions and auroral all sky images. We show that the technique creates embeddings that group together visually similar observations. When provided with novel example images the method correctly identifies similar intervals: when provided with an electron distribution sampled during an encounter with an electron diffusion region the method recovers similar distributions obtained during two other known diffusion region encounters. Similarly, when provided with an interesting auroral structure the method highlights the same structure observed from an adjacent location and at other close time intervals. The method promises to be a useful tool to expand interesting case studies to multiple events, without requiring manual data labeling. Further, the models could be fine-tuned with relatively small set of labeled example data to perform tasks such as classification. The embeddings can also be used as input to deep learning models, providing a key intermediary step—capturing the key features within the data.
AU  - Smith,AW
AU  - Rae,IJ
AU  - Stawarz,JE
AU  - Sun,WJ
AU  - Bentley,S
AU  - Koul,A
DO  - 10.1029/2023JA032096
PY  - 2024///
SN  - 2169-9380
TI  - Automatic Encoding of Unlabeled Two Dimensional Data Enabling Similarity Searches: Electron Diffusion Regions and Auroral Arcs
T2  - Journal of Geophysical Research: Space Physics
UR  - http://dx.doi.org/10.1029/2023JA032096
VL  - 129
ER  -
Download