Guo et al. (2026) Advancing Precipitation Estimation in Mountainous Regions Through Deep Learning Fusion of Multi-Satellite Products

Identification

• Journal: Remote Sensing
• Year: 2026
• Date: 2026-02-15
• Authors: Yinan Guo, Wei Xu, Zhang Zhifu, Jiajia Gao, Li Zhou, Chun Zhou, Lingling Wu, Zhongshun Gu
• DOI: 10.3390/rs18040615

Research Groups

• Xizang Autonomous Region Meteorological Information and Network Centre, China
• National Meteorological Information Center, China
• Xigazê National Climatological Observatory, China Meteorological Administration, China
• School of Ecology and Environment, Xizang University, China
• Institute for Disaster Management and Reconstruction, Sichuan University, China
• State Key Laboratory of Hydraulics and Mountain River Engineering, College of Water Resource & Hydropower, Sichuan University, China
• Sichuan Hydrological and Water Resources Survey Center, China

Short Summary

This study developed a Transformer-based deep learning framework to fuse near-real-time GSMaP-GNRT and IMERG-Early satellite precipitation products, significantly improving precipitation estimation accuracy, particularly bias reduction and monthly statistics, in the mountainous Sichuan Province, China.

Objective

• To develop and evaluate a Transformer-based deep learning fusion framework for near-real-time GSMaP-GNRT and IMERG-Early satellite precipitation products at the gauge scale to generate a high-quality precipitation dataset, supporting improved precipitation estimation, hydrological simulation, and climate change research in complex terrain regions of Sichuan Province.

Study Configuration

• Spatial Scale: Sichuan Province, China. Satellite products at 0.1° × 0.1° (approximately 10 km × 10 km) resolution, evaluated against 156 point-based rain gauges.
• Temporal Scale: 6 years (2015–2020) for data collection and evaluation. Daily and monthly scales for analysis. Transformer model uses a 60-day rolling window for input sequences.

Methodology and Data

• Models used: Transformer deep learning model (encoder-only architecture with 2 encoder layers, 4 attention heads, hidden dimension of 64, and dropout of 0.1). Optimized using Adam to minimize mean squared error (MSE). Station-wise 3-fold cross-validation for evaluation.
• Data sources:
  • Satellite: Global Satellite Mapping of Precipitation (GSMaP-GNRT) and Integrated Multi-satellite Retrievals for GPM (IMERG Early Run) products, both at 0.1° × 0.1° spatial resolution.
  • Observation: Daily precipitation observations from 156 meteorological stations across Sichuan Province (2015–2020) obtained from the National Meteorological Information Center (NMIC) of the China Meteorological Administration (CMA).

Main Results

• All three datasets (GSMaP, IMERG, and the fused product) effectively capture the seasonal precipitation pattern in Sichuan, with higher rainfall in summer and lower in winter.
• At the daily scale, the fused product achieved a Correlation Coefficient (CC) of 0.64, a Bias of 5.21%, and a Root Mean Square Error (RMSE) of 3.83 mm. This represents a reduction in bias relative to both GSMaP (6.24%) and IMERG (-11.46%), and an improvement in RMSE compared to IMERG (4.10 mm), while maintaining a CC comparable to GSMaP (0.72).
• At the monthly scale, the fused product demonstrated the best overall accuracy with the highest CC (0.89), lowest RMSE (44.98 mm), and a Bias of -47.11 mm, outperforming GSMaP (CC=0.83, RMSE=60.64 mm, Bias=-46.59 mm) and IMERG (CC=0.75, RMSE=66.67 mm, Bias=-55.49 mm).
• Spatially, the fused product exhibited reduced bias and RMSE, along with more homogeneous patterns across Sichuan's complex terrain compared to the original satellite products.
• Detection metrics showed that the fused product achieved a higher Probability of Detection (POD, median ~0.85) and a slightly improved Critical Success Index (CSI, median ~0.475). However, the False Alarm Ratio (FAR) remained relatively high (median ~0.50) and comparable to the original products, indicating enhanced event sensitivity and spatial consistency rather than substantially reduced false alarms.

Contributions

• Developed a novel Transformer-based deep learning framework for the station-scale fusion of near-real-time GSMaP-GNRT and IMERG-Early satellite precipitation products, focusing on operational applicability as a satellite-only inference problem.
• Implemented a unique station-wise Transformer that leverages fixed-length rolling windows to effectively capture multi-day temporal dependencies and cross-product interactions between satellite datasets.
• Demonstrated significant improvements in precipitation estimation accuracy, particularly in bias reduction and monthly statistical performance, over complex mountainous terrain in Sichuan Province.
• Enhanced precipitation event detection sensitivity and spatial consistency across the study area, providing a more balanced and reliable precipitation dataset.
• Offered a promising methodological reference and high-quality data foundation for refined precipitation monitoring, hydrological modeling, and disaster risk assessment in mountainous regions.

Funding

• Science and Technology Projects of Xizang Autonomous Region (XZ202501ZY0145)
• Natural Science Foundation Youth Project from Science and Technology Department of Sichuan Province (2024NSFSC0984)

Citation

@article{Guo2026Advancing,
  author = {Guo, Yinan and Xu, Wei and Zhifu, Zhang and Gao, Jiajia and Zhou, Li and Zhou, Chun and Wu, Lingling and Gu, Zhongshun},
  title = {Advancing Precipitation Estimation in Mountainous Regions Through Deep Learning Fusion of Multi-Satellite Products},
  journal = {Remote Sensing},
  year = {2026},
  doi = {10.3390/rs18040615},
  url = {https://doi.org/10.3390/rs18040615}
}