Global MetNet Nowcasting System

• Global MetNet is an operational deep learning-based system for global precipitation nowcasting, integrating satellite, radar, and NWP data.
• Its encoder–decoder architecture employs space-to-depth processing and lead time embedding to deliver high-resolution forecasts in real time.
• The system enhances forecast skill in data-sparse regions and achieves sub-minute inference latency for timely disaster response.

Global MetNet is an operational deep learning-based global weather nowcasting system, designed to provide high-resolution precipitation forecasts for up to 12 hours into the future by leveraging satellite, radar (where available), and numerical weather prediction (NWP) data. It is engineered to overcome the limitations of traditional NWP, especially in regions with sparse ground radar coverage, offering rapid and accurate real-time predictions globally (Agrawal et al., 15 Oct 2025).

1. System Architecture and Design

Global MetNet implements an advanced encoder–decoder neural architecture that builds upon the precedents established by MetNet and MetNet-2. All input sources, including geostationary satellite mosaics, GPM CORRA precipitation estimates, NWP fields (ECMWF HRES and others), and radar (where available), are resampled to a common global resolution of 0.05° × 0.05° (~5 km) and aligned temporally at 15-minute intervals.

The system first applies a space-to-depth operation, whereby fine-scale spatial details are regrouped into additional channels, facilitating efficient representation and preserving meteorologically relevant structures at input resolution. The encoder comprises 4 hierarchical stages, each consisting of stacked residual blocks with skip connections. The number of feature channels increases across stages (from 256, then 384), with cropping operations used after each stage to eliminate border effects due to receptive field growth.

Lead time conditioning is integral to Global MetNet: lead time is discretized (0–12 hours, every 15 minutes) and encoded as a 32-dimensional embedding vector. This vector is injected into the network before each activation, modulating both additively and multiplicatively in FiLM-style layers, ensuring that every internal computation is aware of the desired forecast horizon.

Upsampling is implemented in the decoder to restore full spatial resolution, and multiple output heads produce probabilistic forecasts in the form of categorical Softmax over discretized precipitation bins. Probability thresholds, optimized on validation data to maximize Critical Success Index (CSI), convert these outputs to actionable binary predictions. All training is conducted using the Adam optimizer with Polyak averaging; due to the network’s scale, optimization relies on gradient checkpointing and bfloat16 precision on TPU clusters.

2. Data Sources and Integration

Global MetNet integrates a heterogeneous array of input data:

• GPM CORRA Precipitation: The principal global training target consists of dual-frequency radar and microwave radiometer precipitation measurements from NASA’s GPM CORRA dataset at 0.05° resolution. Despite sparse temporal coverage (satellite revisit ≈2.5 days), CORRA ensures uniform global reach. Patches with missing values are excluded during training, and auxiliary targets may be substituted as necessary.
• Geostationary Satellite Mosaics: Data are collected from 7 satellites (Meteosat, Himawari, GOES, GK-2A), providing multi-band calibrated reflectance and brightness temperature, composited into seamless global mosaics every 15 minutes, using Gaussian blending to mitigate boundary artifacts.
• Global NWP Fields (HRES): Surface and atmospheric variables from ECMWF HRES are resampled to the global grid. Despite higher latency (6–12 hours), these data provide crucial context, especially for longer-range predictions. NWP fields are concatenated as additional channels during encoding.
• Ground Radar: Where available (US, Europe, Japan), high-resolution radar data supplement both input and target, enhancing quality in these regions. Radar observations are strictly used for training and validation within their respective domains.
• Additional Features: Supplementary data include IMERG Early/Final precipitation (use limited by latency), elevation, and latitude/longitude as static predictors.

All inputs are regridded and stacked along the channel dimension, providing the model with temporally and spatially aligned, multi-modal features. This unified input scheme is critical for harmonizing disparate data sources and mitigating spatial and temporal sparsity.

3. Performance Metrics and Validation

Global MetNet’s forecast skill is evaluated with rigorous statistical metrics tailored to precipitation nowcasting:

• Critical Success Index (CSI):

$\operatorname{CSI} = \frac{TP}{TP + FN + FP}$

where TP, FN, FP are true positives, false negatives, and false positives, respectively. Global MetNet exhibits a marked improvement over HRES forecasts across precipitation thresholds (0.2, 1.0, 2.4, 7.0, 25.0 mm/hr), with early-hour CSI gains of up to 0.18 compared to traditional models at low thresholds.

• Fractions Skill Score (FSS): FSS compares spatial patterns between forecast and observation, utilizing varying neighborhood sizes. Global MetNet surpasses baseline NWP methods at both fine (0.05°) and coarser spatial scales.
• Frequency Bias: The system maintains near-unity bias for light precipitation rates, while it purposefully overpredicts (wet bias) at higher rates. This design prioritizes recall for extreme precipitation events, which are operationally critical and often missed by NWP systems.

Validation is performed against both ground radar (where available) and satellite observations, highlighting significant model skill in both data-rich regions and the Global South, where radar is absent.

4. Real-Time Operational Deployment

Global MetNet is engineered for low latency and massive scalability. Total inference latency is typically under one minute, achieved through the elimination of recurrent modules, efficient concatenation of channel-wise temporal slices, and distributed TPU-based execution.

The system is deployed operationally at global scale, notably delivering real-time precipitation forecasts via Google Search to millions of users. The deployment architecture includes optimized thresholding routines and robust input pre-processing pipelines for seamless real-world integration. This rapid turnaround outpaces traditional NWP, whose forecasts often lag by several hours post-initialization.

5. Addressing Global Disparities in Forecast Quality

A principal motivation for Global MetNet is narrowing the persistent gap in weather forecast skill between the Global North and the Global South. By relying chiefly on satellite-based datasets (CORRA, geostationary mosaics) and integrating data without dependence on regional radar coverage, Global MetNet extends high-resolution nowcasting to previously underserved areas. Validation demonstrates that, in tropics and other data-sparse regions, its skill frequently exceeds that of leading NWP models even as measured in the US.

This democratization is critical for vulnerable communities subject to fast-developing storms, providing timely actionable information previously unavailable due to the limitations of both ground infrastructure and conventional NWP latency.

6. Future Research and Methodological Developments

Future work identified by the authors includes refinement of probabilistic outputs, particularly reduction of the wet bias in heavy precipitation categories while retaining high recall. Enhancements in capturing spatial structure and sharpness—especially at extreme rainfall rates—are targeted, with possible integration of additional observational sources (such as lightning detection).

Efforts are underway to broaden accessibility, enabling local meteorological agencies in developing regions to benefit from Global MetNet infrastructure and forecast products. The methodological template suggested—including unified multi-source input, probabilistic output via softmax classification, and lead time embedding—may be generalized for other forms of global impact forecasting.

Summary Table: Global MetNet System Overview

Component	Description	Significance
Encoder–Decoder Design	Deep residual nets, space-to-depth input	Efficient modeling of global context
Data Sources	Satellite mosaics, GPM CORRA, NWP, radar	High coverage in data-sparse regions
Outputs	Probabilistic Softmax, thresholded CSI	Actionable, real-time forecasts
Latency	< 1 minute end-to-end	Operational for disaster response
Validation	CSI, FSS, bias vs. radar/satellite targets	Robust global performance

Global MetNet sets a new standard for operational, globally applicable, deep learning-driven precipitation nowcasting. By effectively integrating disparate data and delivering forecast equity for underserved regions, it marks a significant advance in real-time meteorological prediction (Agrawal et al., 15 Oct 2025).