Huayu-Global: Skillful real-time quasi-global precipitation retrieval

Zijiang Song; Shuang Yang; Yuying Li; Zilong Zhu; Xigang Sun; Ting Liu; Wenye Xiao; Lina Yuan

physics.ao-phOct 21, 2026

Huayu-Global: Skillful real-time quasi-global precipitation retrieval

Authors: Zijiang Song, Shuang Yang, Yuying Li, Zilong Zhu, Xigang Sun, Ting Liu, Wenye Xiao, Lina Yuan,

Organizations: Key Laboratory of Geographic Information Science (Ministry of Education), East China Normal University, Shanghai, 200241, China. · Key Laboratory of Spatial-temporal Big Data Analysis and Application of Natural Resources in Megacities (Ministry of Natural Resources), East China Normal University, Shanghai, 200241, China. · School of Computer Science and Engineering, Southeast University, Nanjing, 211189, China.

Abstract

Satellite precipitation products face a persistent trade-off: real-time estimates are timely but less accurate, whereas the most reliable products are often released months after observation. This delay keeps high-quality precipitation information outside the decision window for flood warning and disaster response. Here we report Huayu-Global, a cross-constellation deep-learning system that combines three geostationary sensor streams from five satellites through a common retrieval framework and generates coherent quasi-global precipitation fields every thirty minutes. Across more than a million independent station records from two periods outside the training years, Huayu-Global improves event detection and precipitation-amount estimation relative to the real-time products evaluated and reaches delayed-product-level skill in key metrics. Huayu-Global therefore turns operational geostationary observations into a timely, high-skill precipitation layer for flood forecasting, water management and disaster response. Its cross-constellation design also provides a scalable foundation for incorporating future geostationary missions into hydrological and weather-prediction systems.

Figures & tables

Figure 1: Data processing workflow for the Huayu-Global precipitation monitoring framework. Representative Geostationary Operational Environmental Satellite (GOES), Meteosat, and FengYun-4B (FY-4B) sector views are transformed into Advanced Baseline Imager (ABI), Spinning Enhanced Visible and Infrared Imager (SEVIRI), and Advanced Geostationary Radiation Imager (AGRI) cloud-feature sequences, then passed through the Huayu models for the Americas, Europe-Africa, and Asia-Pacific domains. Sector outputs are merged to generate coherent quasi-global grid precipitation products using only satellite inputs available at the inference time.

Figure 2: Detection skill by product availability class. The vertical axis shows the full-sample CSI from the independent station evaluation, and the horizontal axis shows documented product-release or processing classes rather than measured end-to-end processing time. Huayu-Global and the real-time baseline products are grouped in the real-time class; IMERG Early Run (IMERG ER), IMERG Late Run (IMERG LR), PERSIANN, CMORPH, and IMERG FR are displayed according to their documented release categories. Lines connect product streams within the same product family where applicable.

Product	Resolution	Latency	CSI $\uparrow$	ACC $\uparrow$	R 2 $\uparrow$	CC $\uparrow$	RMSE $\downarrow$ (mm)
PERSIANN-CCS	0.04°/1h	real-time	$0.234$ $\pm 0.001$	$0.853$ $\pm 0.001$	$-0.177$ $\pm 0.020$	$0.334$ $\pm 0.008$	$2.660$ $\pm 0.032$
GSMaP_NOW	0.1°/1h	real-time	$0.327$ $\pm 0.001$	$0.859$ $\pm 0.001$	$-0.374$ $\pm 0.036$	$0.387$ $\pm 0.008$	$2.874$ $\pm 0.039$
CMORPH	8km/30min	3-4 months	$0.401$ $\pm 0.002$	0.904 $\pm 0.000$	$0.255$ $\pm 0.013$	$0.597$ $\pm 0.008$	$2.118$ $\pm 0.035$
PERSIANN	0.25°/1h	2 days	$0.281$ $\pm 0.001$	$0.855$ $\pm 0.001$	$0.156$ $\pm 0.007$	$0.421$ $\pm 0.007$	$2.257$ $\pm 0.037$
IMERG ER	0.1°/30min	4 hours	$0.404$ $\pm 0.002$	$0.876$ $\pm 0.000$	$0.195$ $\pm 0.013$	$0.553$ $\pm 0.008$	$2.203$ $\pm 0.034$
IMERG LR	0.1°/30min	12 hours	$0.421$ $\pm 0.002$	$0.887$ $\pm 0.000$	$0.244$ $\pm 0.013$	$0.590$ $\pm 0.007$	$2.134$ $\pm 0.034$

Table 1: Overall performance comparison over 1,847,967 matched 3-hour station-product records from July–December 2022 and January–June 2025. Huayu-Global is compared with seven satellite precipitation products on the same matched station windows using CSI, ACC, R 2 , CC, and RMSE. All values in this table are evaluated against independent station observations. Smaller values below each estimate denote

\pm

one-half of the central 95% bootstrap confidence interval from 1,000 station–date-block resamples (Methods). Bold values indicate the best value in each metric column.

Figure 3: Representative spatial comparison of precipitation estimates for one selected evaluation case. Panels show (a) GSMaP_NOW, (b) PERSIANN-CCS, (c) PERSIANN, (d) CMORPH, (e) IMERG FR, and (f) Huayu-Global. Each panel presents a quasi-global precipitation field with four enlarged regional views; station markers indicate the validation locations in the zoomed regions. The annotations report case-specific CSI, CC, R 2 , and RMSE for the displayed case only and should not be interpreted as the full-sample scores in Table 1 .

Region	Records	CSI $\uparrow$	ACC $\uparrow$	R 2 $\uparrow$	RMSE $\downarrow$ (mm)
Americas	302,755	0.495	0.814	0.339	2.452
Europe-Africa	135,130	0.398	0.888	0.247	1.597
Asia-Pacific	1,410,082	0.415	0.914	0.356	1.866
Quasi-global	1,847,967	0.441	0.896	0.362	1.961

Table 2: Regional performance of Huayu-Global across three operational sectors and the quasi-global station sample. Metrics are evaluated over July–December 2022 and January–June 2025. Sector entries are record-count-weighted summaries of independently evaluated longitude segments and are not recomputed from pooled sector samples; the quasi-global entry is computed directly over the full evaluation sample. Because R 2 is computed relative to the range of the records being pooled, and the quasi-global R 2 is computed over the full range of the pooled records, it is not a record-count-weighted average of sector R 2 values; between-sector differences in observed precipitation can therefore make the pooled quasi-global R 2 exceed the individual sector values.

Figure 4: Station-level correlation coefficient (CC) across the Asia validation domain, shown for the station network concentrated in East Asia. Panels a–f show GSMaP_NOW, PERSIANN-CCS, PERSIANN, CMORPH, IMERG FR, and Huayu-Global, respectively. Each circle marks a validation station and is coloured by the Pearson CC between the product estimates and station observations, calculated from all available paired 3-hour records in the July–December 2022 and January–June 2025 validation windows. The colour scale ranges from 0 (cyan) to 1 (magenta).

Figure 5: Stratified performance across validation conditions. Rows show CSI, ACC, POD, FAR, CC, R 2 , and RMSE; columns show validation month, station elevation, longitude sector in 2022 and 2025, and observed 3-hour precipitation amount. Solid teal lines denote Huayu-Global, and dashed grey lines denote the strongest non-Huayu baseline selected separately within each bin and metric. Metric–bin combinations that are undefined or uninformative by construction are omitted.

Extended Data Fig. 1: Architecture of the Huayu-Global retrieval model. Multispectral geostationary imagery from the AGRI aboard FY-4B, the SEVIRI aboard Meteosat, and the ABI aboard GOES is mapped to regional grid precipitation estimates for the Asia-Pacific, Europe-Africa, and Americas domains. Each sector model contains a Head Block, ten Geographic Attention (GeoAB) modules, and a Tail Block, with residual connections supporting feature propagation. The GeoAB module combines local cloud-texture attention with broader geographic context to capture fine-scale convective signatures and large-scale precipitation patterns.

Extended Data Fig. 2: Daily-scale stability of Huayu-Global relative to seven baseline precipitation products. Panel a compares daily CSI, CC, and RMSE trajectories for Huayu-Global, the strongest non-Huayu baseline, and the full baseline range during the July–December 2022 and January–June 2025 validation windows. Panel b summarizes Huayu-Global’s relative daily skill intensity across categorical and continuous metrics. Panel c shows centred 14-day rolling-median differences between Huayu-Global and the strongest non-Huayu baseline, selected separately for each day and metric. Positive vertical displacement denotes stronger Huayu-Global performance, and teal shading marks positive differences. Dashed vertical lines separate the validation periods. Because the strongest baseline is selected separately by metric and date, this figure is a diagnostic envelope comparison rather than a head-to-head comparison with one fixed product. Huayu-Global generally remains near the upper envelope for CSI and CC and near the lower-error envelope for RMSE, although the strongest non-Huayu baseline changes by metric and date.

Extended Data Fig. 3: The same as Fig. 4 but across the Americas.

Extended Data Fig. 4: The same as Fig. 4 but across Europe.

Extended Data Fig. 5: Station-level critical success index (CSI) across the Americas validation domain. Panels a–f show GSMaP_NOW, PERSIANN-CCS, PERSIANN, CMORPH, IMERG FR, and Huayu-Global, respectively. Each circle marks a validation station and is coloured by the CSI for precipitation-event detection, calculated from all available paired 3-hour records in the July–December 2022 and January–June 2025 validation windows using the common precipitation-occurrence definition specified in Methods. The colour scale ranges from 0 (cyan) to 1 (magenta).

Extended Data Fig. 6: The same as Extended Data Fig. 5 but across Europe.

Extended Data Fig. 7: The same as Extended Data Fig. 5 but across Asia.

Parameter	Huayu-ABI	Huayu-AGRI	Huayu-SEVIRI
Satellite	GOES-16/18/19	FY-4B	Meteosat-9/10/11
Learning Rate	0.002	0.001	0.001
Model Parameters	241.78M	241.78M	241.78M
LR Decay Schedule	[10, 20, 30, 40, 50]	[20, 30, 40, 50]	[20, 30, 40, 50]
Training Samples	51,896	23,642	49,691
Checkpoint Epoch	93rd	100th	99th

Extended Data Table 1: Per-sensor training configuration of Huayu-Global. The three sector models share identical architecture and parameter counts, while learning rate and decay schedule are tuned separately for ABI, AGRI, and SEVIRI streams. All models use AdamW optimization, a batch size of 16, and 100 training epochs on four NVIDIA RTX 6000 Ada GPUs (48 GB each). The implementation uses PyTorch with Python 3.12 on Linux. Training data cover 2023 and 2024; model checkpoints are selected according to held-out validation performance.

Satellites	Nominal longitude(s)	Launch date	Payloads	Bands	Native resolution	Provider
FengYun-4B	105°E/133°E	2nd June 2021	AGRI	15	4km/15min	China
GOES-16 GOES-19	75.2°W	19th Nov 2016 25th Jun 2024	ABI	16	2km/10min	U.S.A.
GOES-18	137°W	1st March 2022	ABI	16	2km/10min	U.S.A.
Meteosat-10 Meteosat-11	0°	5th July 2012 15th July 2015	SEVIRI	11	3km/15min	Europe
Meteosat-9	45°E	22nd Dec 2005	SEVIRI	11	3km/15min	Europe

Extended Data Table 2: Operational geostationary satellites used in Huayu-Global. The Americas sector is served by GOES-16/18/19 ABI observations, the Europe-Africa sector by Meteosat-9/10/11 SEVIRI observations, and the Asia-Pacific sector by FY-4B AGRI observations. These three systems together cover most low- and mid-latitude regions (60 ∘ S–60 ∘ N). GOES-16 and GOES-19 are handled as a continuous ABI stream according to provider operational metadata, with GOES-19 contributing after the GOES-East transition in 2025; Meteosat-10 and Meteosat-11 are likewise handled as a continuous SEVIRI stream. The coverage of AGRI (FY-4B) differs before and after 31 January 2024 owing to a change of the sub-satellite point (from 133 ∘ E to 105 ∘ E); the mosaicking procedure accounts for the corresponding sector footprints. The table lists the nominal longitude, launch date, imaging payload, spectral bands, native sampling, and provider for each platform.

Name	Data Size	Format	Provider	Type	Covered Region
GOES-16	9.2 T	grid	NOAA	Full disk	The Americas
GOES-18	11 T	grid	NOAA	Full disk	The Americas
GOES-19	2.4 T	grid	NOAA	Full disk	The Americas
Meteosat-10/11	14 T	grid	EUMETSAT	Full disk	Europe-Africa
Meteosat-9	5.7 T	grid	EUMETSAT	Full disk	Europe-Africa
FY-4B	12 T	grid	NSMC	Full disk	Asia-Pacific

Extended Data Table 3: Data inventory used in this study. The table summarizes raw geostationary observations, baseline precipitation products, independent station validation data, and processed training tensors, with a total rounded volume of 59.7 TB. Processed ABI, SEVIRI, and AGRI tensors are generated by collocating geostationary imagery with IMERG FR reference fields for the three Huayu-Global sector models. Baseline products span real-time (GSMaP_NOW, PERSIANN-CCS), short-delay (IMERG ER, IMERG LR), and delayed post-processed (PERSIANN, CMORPH, IMERG FR) streams. Ground validation uses 5,395 HadISD and CMA stations contributing 1,847,967 three-hourly records from two validation windows (July–December 2022 and January–June 2025) completely outside the 2023–2024 training period.

Remote Sensing

Remote sensing uses satellite, aerial, and ground-based observations to measure and map Earth and other planetary surfaces.

See all 527 papers

A plug-and-play generative framework for multi-satellite precipitation estimation
May 14, 2026Yunfan Yang, Haofei Sun, Xiuyu Sun +7Remote Sensing Generative Modeling
State Key Laboratory of Atmospheric Boundary Layer Physics and Atmospheric Chemistry (LAPC), Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China. · College of Earth and Planetary Sciences, University of Chinese Academy of Sciences, Beijing, China. · Shanghai Academy of Artificial Intelligence for Science (SAIS), Shanghai, China. +5
Neural Stochastic Processes for Satellite Precipitation Refinement
Apr 12, 2026Shunya Nagashima, Takumi Bannai, Shuitsu Koyama +2Neural Processes Remote Sensing
Neurogica Inc. · LTS, Inc. · ME-Lab Japan, Inc.
Station2Radar: query conditioned gaussian splatting for precipitation field
Feb 28, 2026Doyi Kim, Minseok Seo, Changick KimGaussian Splatting Remote Sensing
Korea Advanced Institute of Science and Technology (KAIST)

May 14, 2026J/K move · Enter open · S save

A plug-and-play generative framework for multi-satellite precipitation estimation

Yunfan Yang, Haofei Sun, Xiuyu Sun, Wei Han+6

State Key Laboratory of Atmospheric Boundary Layer Physics and Atmospheric Chemistry (LAPC), Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China. · College of Earth and Planetary Sciences, University of Chinese Academy of Sciences, Beijing, China. · Shanghai Academy of Artificial Intelligence for Science (SAIS), Shanghai, China.+5

Remote Sensing Generative Modeling Multimodal Sensor Fusion

Open paper AskPDF arXiv

Reliable precipitation monitoring is essential for disaster risk reduction, water resources management, and agricultural decision-making. Multi-source satellite observations, particularly the combination of geostationary infrared and passive microwave measurements, have become a primary means of precipitation detection. Traditional multi-source satellite precipitation estimation methods remain computationally inefficient, and many deep learning methods lack the flexibility to incorporate new sensors without retraining the full model. Here we introduce PRISMA (Precipitation Inference from Satellite Modalities via generAtive modeling), a plug-and-play latent generative framework for multi-sensor precipitation estimation. PRISMA learns an unconditional precipitation prior from IMERG Final fields and constrains it through independently trained, sensor-specific conditional branches, allowing new observation sources to be incorporated without retraining the generative backbone. Applied to FY-4B AGRI infrared and GPM GMI microwave observations, PRISMA improves Critical Success Index by up to 40.3% and reduces root-mean-square error by 22.6% relative to infrared-only estimation within microwave swaths, while also improving probabilistic skill and maintaining an average inference time of about 37 s. Independent rain-gauge validation across China confirms consistent gains, and typhoon case studies show that microwave conditioning restores eyewall and spiral rainband structures, reducing storm-core mean absolute error by up to 42.3%. PRISMA thus provides an extensible and efficient framework for multi-sensor precipitation estimation.

Abstract

Remote Sensing

Explore similar work

A plug-and-play generative framework for multi-satellite precipitation estimation

Neural Stochastic Processes for Satellite Precipitation Refinement

Station2Radar: query conditioned gaussian splatting for precipitation field