Skip to main content

Lund University Publications

LUND UNIVERSITY LIBRARIES

Reconstruction of missing radiosonde observations for Indian region using conditional GAN framework

Singh, Gaganpreet ; Durbha, Surya S. ; Budakoti, Sachin LU and C R, Shreelakshmi (2026) In Journal of Atmospheric and Solar-Terrestrial Physics 284.
Abstract

Radiosonde observations represent one of the most reliable and relevant methods of vertical profile measurements of the upper air metrological observations. Despite its importance in Numerical weather prediction (NWP), radiosonde observations suffer from persistent data gaps resulting from sensor failure, critical environment conditions, and signal interference. Therefore, there is a growing need for data reconstruction that may be able to fill the missing records and extend the data usability in NWP models for enhanced short range weather forecasts. Advanced machine learning methods offer a promising path to overcome the problem and increase the reliability of radiosonde observations. Present study utilized a robust approach to... (More)

Radiosonde observations represent one of the most reliable and relevant methods of vertical profile measurements of the upper air metrological observations. Despite its importance in Numerical weather prediction (NWP), radiosonde observations suffer from persistent data gaps resulting from sensor failure, critical environment conditions, and signal interference. Therefore, there is a growing need for data reconstruction that may be able to fill the missing records and extend the data usability in NWP models for enhanced short range weather forecasts. Advanced machine learning methods offer a promising path to overcome the problem and increase the reliability of radiosonde observations. Present study utilized a robust approach to reconstruct missing radiosonde observations across six distinct sites in Indian region during the year 2024 (Delhi, Lucknow, Karaikal, Patna, Pune and Srinagar) using Conditional Generative Adversarial Networks (c-GAN) framework. For robustness and generalization capabilities of AI driven model performance validation metrics (Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Mean Square Error (MSE)) have been considered. From results, it is observed that the coefficient of determination ranges from 0.98 to 0.99 demonstrates that the model performs very well for synthesizing the missing data across all stations. The cGAN model achieves a Mean Absolute Error (MAE) range from 1.5 to 3.0 °C, and RMSE lies between 2.5 °C and 4.5 °C, indicating high reconstruction precision. Present research provides a promising pathway for atmospheric data synthesis based on physics integrated AI method with direct application in a meteorology.

(Less)
Please use this url to cite or link to this publication:
author
; ; and
organization
publishing date
type
Contribution to journal
publication status
published
subject
keywords
c-GAN, Machine learning, Meteorology, Radiosonde, Validation metrics
in
Journal of Atmospheric and Solar-Terrestrial Physics
volume
284
article number
106840
publisher
Elsevier
external identifiers
  • scopus:105039956524
ISSN
1364-6826
DOI
10.1016/j.jastp.2026.106840
language
English
LU publication?
yes
id
9e7481e5-20e1-4b95-877b-6faf02bfebef
date added to LUP
2026-08-11 11:15:28
date last changed
2026-08-12 08:50:13
@article{9e7481e5-20e1-4b95-877b-6faf02bfebef,
  abstract     = {{<p>Radiosonde observations represent one of the most reliable and relevant methods of vertical profile measurements of the upper air metrological observations. Despite its importance in Numerical weather prediction (NWP), radiosonde observations suffer from persistent data gaps resulting from sensor failure, critical environment conditions, and signal interference. Therefore, there is a growing need for data reconstruction that may be able to fill the missing records and extend the data usability in NWP models for enhanced short range weather forecasts. Advanced machine learning methods offer a promising path to overcome the problem and increase the reliability of radiosonde observations. Present study utilized a robust approach to reconstruct missing radiosonde observations across six distinct sites in Indian region during the year 2024 (Delhi, Lucknow, Karaikal, Patna, Pune and Srinagar) using Conditional Generative Adversarial Networks (c-GAN) framework. For robustness and generalization capabilities of AI driven model performance validation metrics (Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Mean Square Error (MSE)) have been considered. From results, it is observed that the coefficient of determination ranges from 0.98 to 0.99 demonstrates that the model performs very well for synthesizing the missing data across all stations. The cGAN model achieves a Mean Absolute Error (MAE) range from 1.5 to 3.0 °C, and RMSE lies between 2.5 °C and 4.5 °C, indicating high reconstruction precision. Present research provides a promising pathway for atmospheric data synthesis based on physics integrated AI method with direct application in a meteorology.</p>}},
  author       = {{Singh, Gaganpreet and Durbha, Surya S. and Budakoti, Sachin and C R, Shreelakshmi}},
  issn         = {{1364-6826}},
  keywords     = {{c-GAN; Machine learning; Meteorology; Radiosonde; Validation metrics}},
  language     = {{eng}},
  publisher    = {{Elsevier}},
  series       = {{Journal of Atmospheric and Solar-Terrestrial Physics}},
  title        = {{Reconstruction of missing radiosonde observations for Indian region using conditional GAN framework}},
  url          = {{http://dx.doi.org/10.1016/j.jastp.2026.106840}},
  doi          = {{10.1016/j.jastp.2026.106840}},
  volume       = {{284}},
  year         = {{2026}},
}