Model Card: GPROF-NN 3D

Model Details

Model Name: GPROF-NN 3D
Developer: Simon Pfreundschuh, Paula J. Brown, Christian D. Kummerow License: MIT Model Type: Neural Network for Precipitation Retrieval
Language: Not applicable Framework: PyTorch Repository: github.com/simonpf/gprof_nn

Model Description

GPROF-NN 3D a precipitation retrieval algorithm for passive microwave (PMW) observations for the sensors of the GPM constellation. It is based on a convolutional neural network leveraging both spatial (2D) and spectral (+1D) information. The version provided here is an early prototype of the model that will become GPROF V8.

Inputs

Brightness temperatures from passive microwave sensors
Earth incidence angles
Ancillary atmospheric and surface state information (e.g., surface temperature, humidity)

Outputs

Surface precipitation estimates
Hydrometeor profiles

Training Data

Training Data Source: Satellite-based observations and collocated ground truth precipitation estimates (e.g., GPM DPR, rain gauges, reanalysis data)
Data Preprocessing: Normalization, quality control, and augmentation techniques applied to enhance generalization

Training Procedure

Optimizer: AdamW
Loss Function: Quantile regression
Training Hardware: 1 A100 GPU
Hyperparameters: Not exhaustively tuned

Performance

Evaluation Metrics: Bias, Mean Squared Error (MSE), Mean Absolute Error (MAE), Correlation Coefficient, Symmetric Mean Absolute Percentage Error (SMAPE)
Benchmark Comparisons: Compared against conventional GPROF algorithm.
Strengths: Lower errors, higher correlation, higher effective resolution
Limitations: Sensitivity to sensor-specific biases

Intended Use

Primary Use Case: Satellite-based precipitation retrieval for weather and climate applications
Potential Applications: Hydrology, extreme weather forecasting, climate research
Usage Recommendations: Performance may vary across different climate regimes

Ethical Considerations

Bias Mitigation: Extensive validation against independent datasets

How to Use

See the external model implementation available from the IPWG ML working group model repository.

Citation

If you use GPROF-NN 3D in your research, please cite:

@Article{amt-17-515-2024,
AUTHOR = {Pfreundschuh, S. and Guilloteau, C. and Brown, P. J. and Kummerow, C. D. and Eriksson, P.},
TITLE = {GPROF V7 and beyond: assessment of current and potential future versions of the GPROF passive microwave precipitation retrievals against ground radar measurements over the continental US and the Pacific Ocean},
JOURNAL = {Atmospheric Measurement Techniques},
VOLUME = {17},
YEAR = {2024},
NUMBER = {2},
PAGES = {515--538},
URL = {https://amt.copernicus.org/articles/17/515/2024/},
DOI = {10.5194/amt-17-515-2024}
}

Contact

For questions see corresponding author in reference.