simonpf/ipwg_net

IPWG-Net: A Baseline CNN model for Satellite-Based Precipitation Estimation

Model Summary

This repository provides trained IPWG-Net models for estimating surface precipitation rates from GMI, geostationary IR, and multi spectral geostationary observations.

Three observation-specific versions are provided:

• ipwg_net_gmi.pt: Uses observations from the GPM Microwave Imager.
• ipwg_net_geo.pt: Uses geostationary visible/infrared observations.
• ipwg_net_geo_ir.pt: Uses infrared-only geostationary observations.

All models use the same basic U-Net architecture with residual convolution blocks. The input scenes have a spatial size of 256 × 256 pixels at 0.036° resolution.

Model Details

Architecture

The baseline model is a U-Net with basic residual blocks. The network follows an encoder–decoder structure with skip connections between corresponding resolution levels. Residual blocks are used throughout the encoder and decoder to improve optimization while keeping the architecture simple and computationally moderate.

The model maps satellite observations to gridded precipitation-rate estimates. The output is a two-dimensional precipitation field with the same spatial dimensions as the input scene.

Inputs

Each model takes a 256 × 256 pixel satellite-observation scene as input. The exact input channels depend on the model variant:

• GMI: 13 channels of passive microwave observations from GMI.
• Geo: 16 channels of geostationary visible/infrared observations.
• Geo-IR: 1 channel of infrared-only geostationary observations.

Input data are expected to follow the preprocessing conventions of the SatRain benchmark dataset, including the dataset-specific normalization, channel ordering, spatial grid, and missing-value handling.

Outputs

The model predicts surface precipitation rate on the SatRain grid.

• Output shape: 256 × 256 pixels
• Spatial resolution: 0.036°
• Output variable: precipitation rate
• Units: mm h⁻¹

Training Data

The models are trained on the SatRain benchmark dataset, which provides collocated satellite observations and reference precipitation estimates. The training scenes consist of satellite observations paired with gauge-corrected ground-radar precipitation estimates.

The benchmark is designed for machine-learning-based precipitation retrieval and provides standardized training, validation, and test splits to enable consistent comparison across retrieval methods.

Intended Use

These models are intended as baseline retrievals for the SatRain benchmark. Suitable uses include:

• Benchmarking new satellite-precipitation retrieval methods.
• Comparing passive microwave and geostationary retrieval skill under a common training and evaluation protocol.

The models are not intended to provide an operational precipitation product without additional validation, calibration, uncertainty characterization, and quality control.

Out-of-Scope Use

The models should not be used as-is for:

• Operational hydrological forecasting.
• Flood warning or emergency response.
• Climate trend analysis.
• Quantitative decision-making in regions or conditions not represented in the SatRain test data.
• Applications requiring well-calibrated uncertainty estimates.

The models provide deterministic precipitation estimates and do not, by themselves, quantify retrieval uncertainty.

Limitations

Satellite-based precipitation retrieval is an ill-posed inverse problem. Retrieval accuracy depends strongly on surface type, storm regime, sensor characteristics, viewing geometry, and the representativeness of the reference data.

Important limitations include:

• Regional generalization: Model performance may degrade outside the geographic and meteorological regimes represented in the training data.
• Sensor dependence: The three model variants are not directly interchangeable because they use different observation types. The ipwg_net_geo model is trained on GOES-16 observations and will fail for application over the Austria testing domain and likely degrade over Korea.
• Heavy precipitation: Extreme precipitation rates will be underestimated due to regression-to-the-mean behavior.
• Light precipitation: Detection of weak or shallow precipitation may be challenging, especially for infrared-only observations.
• Reference uncertainty: Gauge-corrected radar products contain their own uncertainties, including radar artifacts, gauge representativeness errors, and sampling limitations.
• No temporal context: The baseline models operate on individual scenes and do not explicitly use temporal evolution unless such information is encoded in the input observations.
• No uncertainty estimate: The baseline models provide point estimates only.