PaperlayerView exact revision on arXiv

Source-linked AI summary

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, Min Liu

arXiv:2610.29213v1physics.ao-phcs.AI

TL;DR

Timely satellite precipitation estimates often have lower skill than delayed products, while the differing sensors and rainfall-retrieval relationships pose challenges for quasi-global coverage. Huayu-Global combines sensor-adapted geostationary models to generate precipitation fields every 30 minutes; at independent stations, it leads evaluated products on key skill metrics.

  • Problem

    Operational monitoring must balance timely precipitation estimates against quantitative reliability, and cloud-to-rainfall relationships and geostationary sensors vary across regions and platforms.

    Source: §1 · p. 2

  • Method

    Huayu-Global uses a common Head–GeoAB–Tail architecture with separate sensor-stream parameters to retrieve precipitation from ABI, SEVIRI, and AGRI observations.

    Source: §4.2 · p. 12

  • Results

    Huayu-Global achieves the highest CSI and R2 and lowest RMSE among evaluated products at independent stations; relative to GSMaP NOW, it increases CSI by 34.9% and reduces RMSE by 31.8%.

    Source: §3 · p. 10

  • Takeaways & Limitations

    The shared architecture and sensor-specific parameters provide a practical path for incorporating future geostationary missions into a coherent quasi-global product.

    Source: §3 · p. 10

    Coverage is restricted to approximately 60°S–60°N, and validation networks are sparse over Africa, South America, and the oceans.

    Source: §3 · p. 11

Abstract

from arXiv · show

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.

1 Introduction

Operational precipitation monitoring must balance timely delivery with quantitative reliability, while satellite systems face the challenge of retrieving rainfall from differing geostationary sensors. Huayu-Global combines sensor-adapted models across three regions to produce coherent quasi-global fields every 30 minutes, with skill reported against real-time products and delayed-product benchmarks.

1 Introduction

  • Real-time products trade lower detection and quantitative skill for prompt coverage, while the highest-quality streams may arrive months after observation.In the reported evaluation, real-time products had CSI values of 0.23–0.33, compared with 0.42 for IMERG FR, released 3.5 months after observation.Source: §1 · p. 2
  • Cloud-top radiances do not uniquely determine surface rainfall, and retrieval relationships vary across regions, seasons, terrain, storm types, and sensor geometries.GOES ABI, Meteosat SEVIRI, and FY-4B AGRI also differ in spectral coverage, native spatial resolution, and scan cadence.Source: §1 · p. 2
  • Huayu-Global coordinates sensor-adapted GOES ABI, Meteosat SEVIRI, and FY-4B AGRI models across the Americas, Europe-Africa, and Asia-Pacific, merging their outputs on a unified grid.The workflow uses satellite inputs available at inference time to generate coherent quasi-global precipitation fields.Source: Figure 1 · p. 3, §1 · p. 4
  • Huayu-Global leads evaluated real-time products in event detection and error magnitude and reaches delayed-product-level skill in key station-based metrics.It generates fields every 30 minutes and was evaluated at independent stations across two periods outside the 2023–2024 training years.Source: §1 · p. 4

2 Results

Huayu-Global leads the evaluated products on key full-sample detection and error metrics, including delayed products, while a selected case illustrates spatial structure and performance varies across regions and validation conditions.

2.1 Availability–skill comparison

  • Huayu-Global is the only evaluated real-time product with CSI higher than every baseline, including products released months later; it has the highest CSI (0.441) and R2 (0.362) and lowest RMSE (1.961 mm).It exceeds IMERG FR in CSI, R2, and RMSE; ACC (0.896) is close to CMORPH’s 0.904 and CC (0.605) to IMERG FR’s 0.613.Source: Table 1 · p. 5, §2.1 · p. 5

2.2 Representative spatial case

  • In one selected case, Huayu-Global has the highest CSI (0.595), CC (0.686), and R2 (0.455), and the lowest RMSE (1.528 mm) among six displayed products.These case-specific values illustrate rather than establish the productwide ranking; that conclusion rests on the full-sample evaluation.Source: §2.2 · pp. 5–6, Figure 3 · p. 6

2.3 Regional performance

  • Regional performance is heterogeneous across the three sensor sectors.The Americas has the highest CSI (0.495) and RMSE (2.452 mm); Europe–Africa has the lowest RMSE (1.597 mm) and CSI (0.398); Asia-Pacific has the largest sample (1,410,082 records), highest ACC (0.914), and highest sector-level R2 (0.356).Source: §2.3 · p. 7
  • Across station–dataset pairs, CC exceeded CSI in 70.54%, but the metrics are not directly comparable because they differ in definition and sensitivity to precipitation records and errors.Figure 4 maps station-level CC across the East Asia-concentrated validation network; corresponding maps cover the Americas and Europe.Source: §2.3 · p. 7, Figure 4 · p. 8

Station-level CC across Asia

  • CSI and POD generally decline with station elevation while FAR rises; RMSE increases sharply with observed precipitation amount, highlighting weaker high-terrain detection and greater heavy-rainfall magnitude error.The strongest baseline is selected separately within each bin and metric, so it is an upper-envelope diagnostic rather than one fixed product.Source: §2.4 · p. 7, Station-level CC across Asia · p. 8, Figure 5 · p. 9

2.4 Performance across validation conditions

  • Across months and longitude sectors, Huayu-Global generally stays near or above the strongest baseline, although the difference varies and 2025 longitude profiles are more heterogeneous.The 2022 longitude profiles are comparatively stable, and Huayu-Global remains among the leading products across most valid longitude sectors in 2025.Source: §2.4 · p. 7, Figure 5 · p. 9

3 Discussion

Huayu-Global extends regional retrieval into a three-sensor quasi-global system, with the highest CSI and R2 and lowest RMSE among evaluated products. The discussion connects this skill to time-critical environmental workflows while identifying coverage and validation gaps.

3 Discussion

  • CSI rises 34.9% and RMSE falls 31.8% relative to GSMaP NOW; Huayu-Global also leads evaluated products in CSI and R2 and has the lowest RMSE.The system covers approximately 85% of land between 60°S and 60°N.Source: §3 · p. 10
  • A new 0.05° precipitation field is generated every 30 minutes, with loading and inference taking less than 5 minutes per timestep after initialization on one workstation.The fields are framed as a timely observation layer for flood nowcasting, hydrological forecasting, agricultural monitoring, and disaster response.Source: §3 · p. 10
  • Shared architecture with sensor-specific parameters adapts retrievals across ABI, SEVIRI, and AGRI, and offers a path to incorporate future geostationary missions.The named prospective missions are Meteosat Third Generation, FY-4C, and GeoXO.Source: §3 · p. 10
  • Evaluation outside the training years shows aggregate advantage across both windows, while regional results vary: the Americas has CSI 0.495 and RMSE 2.452 mm, and Europe-Africa has CSI 0.398 and RMSE 1.597 mm.The discussion relates regional contrasts to sensor characteristics, climate regimes, and uneven station samples.Source: §3 · p. 10
  • At independent stations, Huayu-Global exceeds its IMERG FR supervision product in CSI, R2, and RMSE, showing that the gridded target need not set the station-scale performance ceiling.Independent radar composites, denser gauge networks, and ocean observations would extend evidence to weakly sampled spatial regimes.Source: §3 · pp. 10–11
  • Geostationary coverage is limited to approximately 60°S–60°N, and land validation is sparse across Africa, South America, and oceans; warm rain, snow, orographic precipitation, and intense convection remain challenging.The stated reason is that cloud-top radiances do not uniquely determine surface rainfall.Source: §3 · p. 11
  • Elevation-dependent degradation and intensity-dependent RMSE growth motivate terrain and orographic constraints, loss reweighting, and passive-microwave fusion for heavy rainfall.The discussion also identifies ensemble uncertainty, robustness to missing scans, and polar-orbiting microwave observations as development priorities.Source: §3 · p. 11

4 Methods

Huayu-Global harmonizes three geostationary sensor streams, trains sensor-specific models within a shared architecture, and evaluates their precipitation estimates against independent station observations. The methods specify preprocessing, training, evaluation, uncertainty estimation, and the scope of the source data.

4.1 Preprocessing

  • Preprocessing reprojects multispectral observations to a common 0.05° grid, standardizes channels separately by sensor, and interpolates invalid pixels unless gaps persist across the interpolation window.Persistent-gap regions are excluded from training and evaluation; sector outputs are mosaicked with averaging across overlaps.Source: §4.1 · p. 11

4.2 Model framework

  • Three independently trained sensor models share a Head–GeoAB–Tail architecture, then merge sector outputs into a continuous quasi-global product.Separate weights adapt to sensor-specific spectral responses, viewing geometry, and scan cadence while retaining a common framework.Source: §4.2 · p. 12

4.3 Runtime performance

  • On the specified workstation, a subsequent timestamp required 290.2 s for data loading and inference, excluding upstream delivery and downstream dissemination.The first run took 381.7 s including model initialization and weight loading.Source: §4.3 · p. 12

4.4 Training and optimization

  • Each 30-minute output is trained against a collocated IMERG FR field, with separate occurrence and amount terms and regression loss applied only to positive-rainfall pixels.At inference, pixels classified as non-raining are set to zero; raining pixels retain the regression estimate.Source: §4.4 · pp. 12–13

4.5 Evaluation metrics

  • Event detection is evaluated with CSI, POD, FAR, and ACC, while quantitative consistency is assessed with CC, R2, and RMSE.The evaluation uses a 0.1 mm precipitation threshold for occurrence and intensity comparisons.Source: §4.5 · p. 14

4.6 Statistics and uncertainty

  • Bootstrap uncertainty uses 1,000 resamples of station-and-UTC-period blocks to estimate central 95% intervals.The intervals represent sampling uncertainty over the matched station network, not structural uncertainty or unobserved oceanic regions.Source: §4.6 · p. 14

4.7 Data

  • Evaluation uses independent gauge comparisons across 5,395 stations and 1,847,967 three-hour records, with IMERG FR used for training rather than independent validation.The station network is unevenly distributed and does not directly evaluate open-ocean precipitation.Source: §4.7 · p. 15

Data availability

Satellite and comparison-product data are available through their originating providers, subject to access terms and restrictions.

Data access

  • GOES ABI, Meteosat SEVIRI, and FY-4B AGRI observations are available from NOAA NCEI, EUMETSAT, and NSMC, respectively; provider registration or data-use terms may apply.CMA station observations cannot be redistributed, and processed training tensors and matched validation records are not publicly redistributed because of volume and licensing conditions.Source: Data availability · p. 20
  • The trained model files, normalization data, coverage masks, and demonstration observations are available from the Huayu-Global repository.The repository also links to the original archives for comparison products and HadISD observations.Source: Data availability · p. 20

Code availability

The Huayu-Global inference and multisatellite mosaicking code is publicly available under the Apache License 2.0.

Code availability

  • The repository provides setup instructions, inference and mosaicking scripts, trained model links, and demonstration input data for reproducing the public inference workflow.Training code paths, restricted CMA station data, and the complete station-validation workflow are not publicly redistributed.Source: Code availability · p. 20

Extended Data

The extended data materials include a model-architecture description and labels for daily metric comparisons.

Architecture

  • Each sector model comprises a Head Block, ten Geographic Attention modules, and a Tail Block, with residual connections; the modules combine local cloud-texture attention and broader geographic context.The architecture maps multispectral imagery from three geostationary sensor streams to regional precipitation estimates.Source: Extended Data · p. 22

Daily comparisons

  • The daily-comparison labels distinguish metric trajectories, skill intensity, Huayu-Global, the best non-Huayu baseline, and the baseline range.The provided labels do not state the outcome of those comparisons.Source: Extended Data · p. 23

RMSE (mm)

Huayu-Global generally maintains strong daily skill across validation windows, with station-level comparisons and regional maps documenting its performance and evaluation scope.

RMSE (mm)

  • Huayu-Global generally remains near the upper envelope for daily CSI and CC and near the lower-error envelope for RMSE, against a changing baseline.The comparison is diagnostic, not head-to-head with one fixed product, because the strongest baseline changes by metric and date.Source: RMSE (mm) · p. 23

Station-level CC across parts of the Americas

Station-level CC across Europe

Station-level CSI across parts of the Americas

  • Extended Data Fig. 5 maps station-level CSI across the Americas for six precipitation products, using paired 3-hour records from both validation windows.CSI measures precipitation-event detection; station colors span 0 to 1.Source: Station-level CSI across parts of the… · p. 26

Station-level CSI across Europe

Station-level CSI across Asia

  • Extended Data Fig. 7 presents the same station-level CSI display as Extended Data Fig. 5, but across Asia.Source: Station-level CSI across Asia · p. 28
  • Three operational geostationary systems serve the Americas, Europe–Africa, and Asia-Pacific, together covering most regions from 60°S to 60°N.The systems use ABI, SEVIRI, and AGRI observations, respectively.Source: Station-level CSI across Asia · p. 29
  • Validation used 5,395 stations and 1,847,967 three-hour records from July–December 2022 and January–June 2025, outside the 2023–2024 training period.The inventory totals 59.7 TB and includes geostationary observations, baseline precipitation products, station data, and processed training tensors.Source: Station-level CSI across Asia · p. 30

Related papers

Loading 2610.29213v1…