Imputing Missing Long-Term Spatiotemporal Multivariate Atmospheric Data with CNN-Transformer Machine Learning
Fuente:
arXiv
Salvato in:
| Autori principali: | , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866918133595373568 |
|---|---|
| author | Hu, Jiahui Dong, Wenjun Liu, Alan Z. |
| author_facet | Hu, Jiahui Dong, Wenjun Liu, Alan Z. |
| contents | Continuous physical domains are important for scientific investigations of dynamical processes in the atmosphere. However, missing data arising from operational constraints and adverse environmental conditions pose significant challenges to accurate analysis and modeling. To address this limitation, we propose a novel hybrid Convolutional Neural Network (CNN) Transformer machine learning model for multivariable atmospheric data imputation, termed CT-MVP. This framework integrates CNNs for local feature extraction with transformers for capturing long-range dependencies across time and altitude. The model is trained and evaluated on a testbed using the Specified Dynamics Whole Atmosphere Community Climate Model with thermosphere and ionosphere extension (SD-WACCM-X) dataset spanning 13 years, which provides continuous global coverage of atmospheric variables, including temperature and zonal and meridional winds. This setup ensures that the ML approach can be rigorously assessed under diverse data-gap conditions. The hybrid framework enables effective reconstruction of missing values in high-dimensional atmospheric datasets, with comparative evaluations against traditional methods and a simple transformer. The results demonstrate that CT-MVP achieves superior performance compared with traditional approaches, particularly in cases involving extended periods of missing data, and slightly outperforms a simple transformer with the same hyper-parameters. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2509_01141 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Imputing Missing Long-Term Spatiotemporal Multivariate Atmospheric Data with CNN-Transformer Machine Learning Hu, Jiahui Dong, Wenjun Liu, Alan Z. Atmospheric and Oceanic Physics Continuous physical domains are important for scientific investigations of dynamical processes in the atmosphere. However, missing data arising from operational constraints and adverse environmental conditions pose significant challenges to accurate analysis and modeling. To address this limitation, we propose a novel hybrid Convolutional Neural Network (CNN) Transformer machine learning model for multivariable atmospheric data imputation, termed CT-MVP. This framework integrates CNNs for local feature extraction with transformers for capturing long-range dependencies across time and altitude. The model is trained and evaluated on a testbed using the Specified Dynamics Whole Atmosphere Community Climate Model with thermosphere and ionosphere extension (SD-WACCM-X) dataset spanning 13 years, which provides continuous global coverage of atmospheric variables, including temperature and zonal and meridional winds. This setup ensures that the ML approach can be rigorously assessed under diverse data-gap conditions. The hybrid framework enables effective reconstruction of missing values in high-dimensional atmospheric datasets, with comparative evaluations against traditional methods and a simple transformer. The results demonstrate that CT-MVP achieves superior performance compared with traditional approaches, particularly in cases involving extended periods of missing data, and slightly outperforms a simple transformer with the same hyper-parameters. |
| title | Imputing Missing Long-Term Spatiotemporal Multivariate Atmospheric Data with CNN-Transformer Machine Learning |
| topic | Atmospheric and Oceanic Physics |
| url | https://arxiv.org/abs/2509.01141 |