Huang et al. (2026) Deep Learning-Based Multi-Source Precipitation Fusion and Its Utility for Hydrological Simulation

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Atmosphere
• Year: 2026
• Date: 2026-01-08
• Authors: Zihao Huang, Changbo Jiang, Yuannan Long, Shixiong Yan, Yue Qi, Munan Xu, Tao Xiang
• DOI: 10.3390/atmos17010070

Research Groups

Not explicitly mentioned in the provided paper text.

Short Summary

This study develops an attention-enhanced CNN–LSTM framework to correct biases in GPM IMERG V07 precipitation data in a gauge-sparse mountainous basin, demonstrating that correcting a single high-quality satellite product with heterogeneous covariates significantly improves precipitation estimates and hydrological utility compared to multi-source aggregation.

Objective

• To develop and evaluate an attention-enhanced CNN–LSTM (A-CNN–LSTM) bias-correction framework for the GPM IMERG V07 satellite precipitation product in a gauge-sparse mountainous basin, and to assess its impact on precipitation estimates and subsequent hydrological simulations.

Study Configuration

• Spatial Scale: Lüshui River basin (mountainous, gauge-sparse).
• Temporal Scale: Daily.

Methodology and Data

• Models used: Attention-enhanced CNN–LSTM (A-CNN–LSTM) for bias correction, CNN–LSTM, LSTM (for comparison), SWAT (hydrological model).
• Data sources: GPM IMERG V07 (primary microwave satellite precipitation), CHIRPS (satellite precipitation), ERA5 evaporation (reanalysis), Digital Elevation Model (DEM), gauge-interpolated precipitation (for validation).

Main Results

• The IMERG-driven A-CNN–LSTM framework significantly reduced daily root-mean-square error (RMSE) and improved the intensity and timing of 10–50 mm·d⁻¹ rainfall events.
• A single-source IMERG configuration with A-CNN–LSTM outperformed CHIRPS-including multi-source setups in terms of correlation, RMSE, and performance across rainfall-intensity classes.
• When corrected IMERG forced the SWAT model, daily Nash-Sutcliffe Efficiency (NSE) increased from approximately 0.71/0.70 to 0.85/0.79 in calibration/validation periods, respectively.
• SWAT runoff simulation RMSE decreased from 87.92 m³ s⁻¹ to 60.98 m³ s⁻¹.
• Simulated flood peaks and timing closely matched those driven by gauge-interpolated precipitation.

Contributions

• Demonstrates the effectiveness of correcting a single high-quality, widely validated microwave satellite product (IMERG) using a small set of heterogeneous covariates (ERA5 evaporation, DEM) with an A-CNN–LSTM framework in gauge-sparse mountainous basins.
• Provides evidence that this targeted correction strategy is more effective for improving precipitation inputs and their hydrological utility than simply aggregating multiple same-type satellite products.
• Introduces and validates an A-CNN–LSTM framework for satellite precipitation bias correction, showing its superior performance over simpler network architectures (LSTM, CNN–LSTM).

Funding

Not explicitly mentioned in the provided paper text.

Citation

@article{Huang2026Deep,
  author = {Huang, Zihao and Jiang, Changbo and Long, Yuannan and Yan, Shixiong and Qi, Yue and Xu, Munan and Xiang, Tao},
  title = {Deep Learning-Based Multi-Source Precipitation Fusion and Its Utility for Hydrological Simulation},
  journal = {Atmosphere},
  year = {2026},
  doi = {10.3390/atmos17010070},
  url = {https://doi.org/10.3390/atmos17010070}
}