PARDIG: Global Precipitation Estimates Learned from Sparse Satellite Observations
Abstract
Estimating precipitation at the global scale is vital for Earth system science and disaster response, yet observational coverage and measurement precision remain constrained. Existing solutions provide options that trade off desirable traits such as skill, latency, geographical coverage, temporal coverage, and spatial and temporal resolution. We present PARDIG, an artificial intelligence framework based on a spherical graph-mesh transformer architecture that produces near-real-time global precipitation estimates. PARDIG is trained on dense multi-modal inputs, including atmospheric state analyses and multi-spectral geostationary satellite imagery, to predict sparse, high-precision precipitation measurements from spaceborne radar. At inference, the trained model is applied densely across the globe at $0.1^\circ$ spatial and 15-minute temporal resolution, producing complete near-real-time precipitation fields from 2016 to the present, alongside a historical variant extending the record back to 1999. Evaluation against three independent radar networks (over the US, Europe, and Japan) and against over 36,000 surface stations worldwide shows that PARDIG consistently outperforms IMERG V07 (Final and Early), MSWEP V3, ERA5, and HRES across precipitation thresholds and regions, with the largest improvements at extreme intensities. Case studies of extreme precipitation events demonstrate that PARDIG successfully captures their spatial structure and timing. The improvements also extend to forecasting: WeatherNext 3, a global weather model, produces more skillful precipitation forecasts when trained to predict PARDIG. Both the historical and near-real-time PARDIG datasets are openly available under a CC BY 4.0 license.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
PARDIG introduces a sparse-to-dense learning paradigm for global precipitation estimation, leveraging a graph-mesh transformer to map multi-modal satellite and atmospheric inputs to high-precision spaceborne radar targets. The paper demonstrates state-of-the-art performance across diverse regions and precipitation intensities, with rigorous cross-modal validation and open data release, significantly advancing the field of AI-driven Earth system science.
Methodology
The paper proposes PARDIG, a framework that adapts the spherical graph-mesh transformer architecture from WeatherNext 3 for single-pass probabilistic precipitation retrieval. The core methodological innovation is the use of a continuous coordinate-query output head to map dense multi-modal inputs (atmospheric state analyses and geostationary satellite imagery) to sparse, high-precision spaceborne radar (GPM) observations. By training on sparse GPM CORRA footprints and inferring densely at 0.1° resolution, the model avoids the spatial gaps inherent in satellite-only products and the physical biases of reanalysis precipitation. The probabilistic output (binned classification) is a robust choice for skewed precipitation distributions, avoiding the conditional-mean bias of regression losses. The exclusion of ground-based data from training ensures that the model's skill is not tied to local gauge density, allowing for true global generalization.
Results & evaluation
The evaluation is exceptionally rigorous and comprehensive. The authors benchmark against three independent regional radar networks (US, Europe, Japan) and over 36,000 global surface stations, ensuring cross-modal validation. The results show consistent outperformance of state-of-the-art products (IMERG, MSWEP, ERA5) across all precipitation thresholds, with particularly significant gains at extreme intensities (e.g., CSI of 0.12 vs 0.03 at 25 mm/h). The inclusion of case studies for extreme events (flash floods, typhoons) and the demonstration of improved downstream forecasting skill when using PARDIG as a training target for WeatherNext 3 further solidify the technical impact. The evaluation also includes spectral analysis and diurnal cycle checks, confirming physical consistency.
Reproducibility
The paper provides high reproducibility. The datasets (PARDIG NRT and LR) are openly available under a CC BY 4.0 license on Google Cloud Storage and Google Earth Engine. The architecture is based on the publicly described WeatherNext 3 design, and the authors provide detailed specifications of input channels, training targets, and inference procedures. While the full codebase may not be open-sourced in the traditional sense, the release of the large-scale datasets and the clear description of the pipeline allow for significant reproduction and downstream research.
Limitations
The model's skill at polar latitudes (above 67°N/S) is limited due to the lack of GPM coverage in those regions, relying on unconstrained extrapolation. The paper acknowledges a distributional shift in the historical variant (LR) prior to 2008 due to sparser early satellite observations. Additionally, the model shows some difficulty with complex orographic precipitation in high-elevation mountainous sites, potentially due to limitations in the GPM radar target for light snowfall.
Broader impact
PARDIG has significant broader impact by providing a high-resolution, near-real-time global precipitation dataset that is independent of ground-based infrastructure. This is crucial for disaster response, hydrological modeling, and climate adaptation in sparsely gauged regions and oceans. The open release of the data under a permissive license will likely accelerate research in Earth system science and improve the accuracy of global weather models, as demonstrated by the improved forecasting skill in WeatherNext 3. PARDIG introduces a sparse-to-dense learning paradigm for global precipitation estimation, leveraging a graph-mesh transformer to map multi-modal satellite and atmospheric inputs to high-precision spaceborne radar targets. The paper demonstrates state-of-the-art performance across diverse regions and precipitation intensities, with rigorous cross-modal validation and open data release, significantly advancing the field of AI-driven Earth system science.
Author of this paper? Send a correction.