Liu et al. (2026) Precipitation nowcasting based on convolutional LSTM with spatio-temporal information transformation using multi-meteorological factors

Identification

• Journal: Geoscientific model development
• Year: 2026
• Date: 2026-08-03
• Authors: Dufu Liu, Feihu Huang, Peng Zheng, Xiaomeng Huang, Xi Wu, Yuan Xia, Jiafeng Zheng, Xiaojie Li, Jing Fang Hu
• DOI: 10.5194/gmd-19-7089-2026

Research Groups

• School of Computer, Chengdu University of Information Technology
• Department of Earth System Science, Ministry of Education Key Laboratory for Earth System Modelling Institute for Global Change Studies, Tsinghua University

Short Summary

This study proposes a novel dual encoder-decoder neural network (STI-DEDN) that enables the approximation of Spatio-Temporal Information Transformation (STI) equations and robust nonlinear mapping of high-dimensional meteorological and VIL precipitation variables. The model achieves superior predictive accuracy and reduced error rates in hourly multi-step forecasting compared to state-of-the-art deep learning benchmarks.

Objective

• To develop a novel framework for precipitation nowcasting based on convolutional LSTM with spatio-temporal information transformation using multiple-meteorological factors.
• To address the limitations of existing STI equation solvers and short-term data in forecasting short-duration heavy precipitation events.

Study Configuration

• Spatial Scale: High-resolution radar image data captured by the GOES-16 geostationary satellite and the NEXRAD weather radar, with a spatial resolution of 1 km × 1 km.
• Temporal Scale: Temporal resolution of 5 min for precipitation data, covering approximately 10,000 weather events across the continental United States from 2017 to 2019.

Methodology and Data

• Models used: STI-DEDN (Spatio-Temporal Information-Dual Encoder-Decoder Network)
• Data sources: SEVIR (Storm Event Imagery) dataset, which includes four satellite sensor types and covers five distinct data categories.

Main Results

• The proposed STI-DEDN model achieves superior predictive accuracy and reduced error rates in hourly multi-step forecasting compared to state-of-the-art deep learning benchmarks.
• The model effectively captures the spatio-temporal dependencies between heterogeneous meteorological variables and VIL precipitation patterns.
• The adaptive weighted gradient loss (ADGLoss) function is proposed to mitigate the prediction ambiguity caused by the extension of prediction time and rectify systematic underestimation of high-intensity precipitation regions.

Contributions

• This study proposes a novel dual encoder-decoder neural network that enables the approximation of STI equations and robust nonlinear mapping of high-dimensional meteorological and VIL precipitation variables.
• The model addresses the limitations of existing STI equation solvers and short-term data in forecasting short-duration heavy precipitation events.
• The adaptive weighted gradient loss function is proposed to mitigate prediction ambiguity and underestimation biases in precipitation intensity.

Funding

• This research was funded by the National Natural Science Foundation of China (Grant No. 62176181) and the Ministry of Education Key Laboratory for Earth System Modelling Institute for Global Change Studies, Tsinghua University.

Citation

@article{Liu2026Precipitation,
  author = {Liu, Dufu and Huang, Feihu and Zheng, Peng and Huang, Xiaomeng and Wu, Xi and Xia, Yuan and Zheng, Jiafeng and Li, Xiaojie and Hu, Jing Fang},
  title = {Precipitation nowcasting based on convolutional LSTM with spatio-temporal information transformation using multi-meteorological factors},
  journal = {Geoscientific model development},
  year = {2026},
  doi = {10.5194/gmd-19-7089-2026},
  url = {https://doi.org/10.5194/gmd-19-7089-2026}
}