Zhang et al. (2026) Three-dimensional cloud radar reflectivity reconstruction from geostationary multispectral imagery using a context-aware Transformer

Identification

• Journal: International Journal of Applied Earth Observation and Geoinformation
• Year: 2026
• Date: 2026-04-03
• Authors: Shenglan Zhang, Shihao Zhang, Ying Zhou, Hailei Liu, Minzheng Duan
• DOI: 10.1016/j.jag.2026.105275

Research Groups

• Key Laboratory of Atmospheric Sounding, Chengdu University of Information Technology, Chengdu 610225, China
• Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing 100029, China

Short Summary

This study develops a novel Transformer-based framework to retrieve continuous three-dimensional (3D) radar reflectivity fields from geostationary satellite imagery, demonstrating robust performance against CloudSat observations (R=0.80, RMSE=6.75 dBZ for composite reflectivity) and utility in monitoring severe weather like hurricanes.

Objective

• To develop and validate a Transformer-based machine learning framework capable of estimating continuous three-dimensional (3D) cloud radar reflectivity structures from two-dimensional passive geostationary multispectral imagery.

Study Configuration

• Spatial Scale: Continental United States (CONUS), Western Hemisphere (GOES-16 full-disk), four major Atlantic hurricanes (Laura, Paulette, Delta, Zeta), and three ground-based ARM sites (Houston, Southern Great Plains, Gunnison-Crested Butte Regional Airport). Vertical profiles span 64 layers with 240 m spacing, covering an altitude range of approximately 0–15.4 km.
• Temporal Scale: Collocated ABI–CloudSat pairs from 2019–2020 for training and validation. Independent validation with ground-based radar in March 2022. Hurricane analysis for 2020 events. Geostationary satellite observations provide data at 10-minute intervals, enabling continuous day-and-night retrieval.

Methodology and Data

• Models used:
  • Transformer-based deep learning framework.
  • Two-stage retrieval system: a classification Transformer for cloud mask detection and a regression Transformer for radar reflectivity estimation in cloudy layers.
  • Context-aware Transformer utilizing a sliding window of 32 adjacent satellite footprints.
  • Rotary Positional Embeddings (RoPE) for encoding relative ordering within the window.
• Data sources:
  • Geostationary Satellite Imagery: GOES-R Advanced Baseline Imager (ABI) Level-1b radiances (Channels 1, 3–16).
  • Spaceborne Cloud Radar: CloudSat 2B-GEOPROF product (radar reflectivity profiles, 64 vertical bins, 240 m spacing).
  • Ground-based Cloud Radar: Ka-band ARM Zenith Radar (KAZR) observations from three ARM sites (HOU, SGP, GUC) for independent structural validation.
  • Tropical Cyclone Best-Track Data: International Best Track Archive for Climate Stewardship (IBTrACS) for hurricane analysis.

Main Results

• The proposed retrieval framework achieves robust performance for composite reflectivity, with a Pearson correlation coefficient (R) of 0.80 and a root-mean-square error (RMSE) of 6.75 dBZ for the Vis + IR model.
• For layer-wise reflectivity, the Vis + IR model yields R = 0.72 and RMSE = 8.15 dBZ, outperforming the IR-only model (R = 0.66, RMSE = 8.75 dBZ). Both models exhibit negligible mean biases.
• Cloud detection accuracy exceeds 80% at most altitudes, with a Probability of Detection (POD) reaching 70% and a False Alarm Rate (FAR) around 20% between 2 km and 10 km altitude.
• The context-aware Transformer (32-footprint window) significantly improves performance compared to single-pixel input, demonstrating the value of spatial context.
• Retrieval accuracy is highest in the upper troposphere (>6 km), consistent with the satellite's top-down viewing geometry.
• Independent validation against ground-based Ka-band ARM Zenith Radar (KAZR) observations shows high structural similarity (SSIM > 0.872), confirming the model's ability to reproduce realistic vertical cloud morphology across diverse climate regimes.
• The framework successfully captures the three-dimensional evolution of major Atlantic hurricanes, including eyewall consolidation and intensification processes.
• Retrieved composite reflectivity metrics exhibit a strong correlation (R > 0.8) with maximum sustained wind speeds for the analyzed hurricanes (e.g., Laura R=0.87).

Contributions

• Introduces a novel Transformer-based deep learning framework for continuous, high-resolution 3D cloud radar reflectivity reconstruction from geostationary multispectral imagery.
• Addresses the limitations of sparse coverage and narrow swaths of active spaceborne radars by providing wide-field, high-frequency 3D cloud structure information.
• Demonstrates the effectiveness of a context-aware Transformer in leveraging mesoscale cloud organization for improved retrieval accuracy compared to pixel-wise methods.
• Enables continuous day-and-night retrieval of 3D cloud structures, critical for monitoring rapidly evolving weather systems.
• Provides a quantitative basis for severe weather analysis, particularly for hurricane monitoring and intensity estimation, by linking retrieved 3D cloud fields to storm evolution and maximum sustained wind speeds.
• Offers valuable complementary information to sparse active profiling observations, bridging the gap between high temporal resolution imagers and vertical profiling radars.

Funding

• National Natural Science Foundation of China (Grant 42030107)
• National Key R&D Program of China (Grant 2021YFC3090203)

Citation

@article{Zhang2026Threedimensional,
  author = {Zhang, Shenglan and Zhang, Shihao and Zhou, Ying and Liu, Hailei and Duan, Minzheng},
  title = {Three-dimensional cloud radar reflectivity reconstruction from geostationary multispectral imagery using a context-aware Transformer},
  journal = {International Journal of Applied Earth Observation and Geoinformation},
  year = {2026},
  doi = {10.1016/j.jag.2026.105275},
  url = {https://doi.org/10.1016/j.jag.2026.105275}
}