Beyond Ensemble Averages: Leveraging Climate Model Ensembles for Subseasonal Forecasting

This repository contains the supplementary code for our paper on leveraging climate model ensembles for improved subseasonal forecasting performance beyond traditional ensemble averaging approaches.

📄 Paper

Beyond Ensemble Averages: Leveraging Climate Model Ensembles for Subseasonal Forecasting
Artificial Intelligence for the Earth Systems, Vol. 3, No. 4 (2024)

🗂️ Data Description

The dataset includes comprehensive climate model outputs and observational data:

Data Sources

• NCEP-NSCv2 ensemble members
• NASA-GMAO ensemble members
• Ground truth data: precipitation and 2-meter temperature
• Observational data: sea level pressure (slp), relative humidity (rhum), 500mb geopotential height (hgt500)
• Principal components of sea surface temperatures (SSTs)

Additional Files

• US mask for regional analysis
• Observational climatology
• Model climatology
• 33rd and 66th percentile threshold values

Data Structure

All data has been preprocessed and regridded to ensure consistency across ensemble members, ground truth, and climate variables.

train_val/     # Training and validation data
test/          # Test data

Data Dimensions: (t, 64, 128) where:

• Training/validation: t = 312
• NCEP test data: t = 117
• NASA test data: t = 85

Data Access

Download the complete dataset: SSF Data

Prerequisites

System Requirements

• Python 3.3 or higher
• Additional dependencies listed in requirements.txt
• Segmentation Models PyTorch (smp)

It can be installed as

pip install -r requirements.txt

Getting Started

Data Utilities

Use utils_data.py for data loading and preprocessing functions.

Example Notebooks

We provide several Jupyter notebooks demonstrating different modeling approaches:

1. Regression and Tercile Classification

• Regression Baselines + Temperature Tercile Classification

  • Random Forest regression with tercile classification for temperature

• U-Net Regression + Temperature Tercile Classification

  • Deep learning approach using U-Net architecture

2. Advanced Modeling Techniques

• Model Stacking

  • Ensemble methods combining multiple model predictions

• Precipitation Tercile Classification

  • Categorical forecasting for precipitation events

🔬 Methodology

Our approach goes beyond simple ensemble averaging by:

1. Leveraging individual ensemble members rather than just ensemble means
2. Using deep learning to capture spatial patterns and relationships
3. Implementing tercile-based classification for categorical forecasting
4. Combining multiple models through stacking techniques

📈 Results

The methods demonstrated in this repository show improved subseasonal forecasting performance compared to traditional ensemble averaging, particularly for extreme weather events and spatial pattern prediction.

📚 Citation

@article{orlova2024beyond,
  title={Beyond ensemble averages: Leveraging climate model ensembles for subseasonal forecasting},
  author={Orlova, Elena and Liu, Haokun and Rossellini, Raphael and Cash, Benjamin A and Willett, Rebecca},
  journal={Artificial Intelligence for the Earth Systems},
  volume={3},
  number={4},
  pages={e230103},
  year={2024},
  publisher={American Meteorological Society}
}