🌍 Self-Supervised Learning with E-OBS Climate Data

This repository demonstrates self-supervised learning approaches for E-OBS precipitation data from the Copernicus Climate Change Service.

🌦️ Dataset

We work with E-OBS daily precipitation data from the European Climate Assessment & Dataset project. The dataset provides:

• Daily precipitation sums across Europe
• 0.25° grid resolution (~25km spatial resolution)
• Time period: 1950-2024 (version 31.0e)
• Spatial coverage: Europe (-40.4°E to 75.4°E, 25.4°N to 75.4°N)

🚀 Quick Start Dataset

For faster development and testing, we provide a subsampled dataset with 2000 time steps (1950-1955) that is loaded by default:

• Subsampled files: rr_ens_mean_0.25deg_reg_v31.0e_sub2000.nc_sub (~27MB each)
• Time range: 1950-01-01 to 1955-06-23 (5.5 years)
• Original files: rr_ens_mean_0.25deg_reg_v31.0e.nc (~348MB each)
• Automatic fallback: If subsampled files are not found, the full dataset is loaded

💡 Note: The data loader automatically prefers the subsampled files for faster loading and experimentation. To force loading the full dataset, use loader.load_precipitation_data(prefer_subsampled=False).

🎯 Self-Supervised Learning Approaches

1. Temporal Prediction 🔮

• Task: Predict next day's precipitation from past 3 days
• Architecture: Transformer-based model with patch embeddings
• Learning: Temporal patterns and weather evolution
• Use case: Weather forecasting, temporal pattern learning

2. Masked Modeling 🎭

• Task: Reconstruct masked spatial regions of precipitation maps
• Architecture: U-Net encoder-decoder with skip connections
• Masking strategies: Random patches, blocks, irregular patterns
• Use case: Data imputation, spatial relationship learning

🚀 Quick Start

Installation

pip install -r requirements.txt

Run the Example

Open and run the main notebook:

jupyter notebook notebooks/self_supervised_eobs_example.ipynb

This notebook demonstrates:

• Loading and exploring E-OBS precipitation data
• Creating temporal prediction datasets
• Training temporal prediction models
• Creating masked modeling datasets
• Training masked modeling models
• Evaluating and visualizing results

📁 Repository Structure

hydrology_seminar/
├── notebooks/
│   └── self_supervised_eobs_example.ipynb    # Main demonstration notebook
├── src/
|   ├── data/
|   |   ├── rr_ens_mean_0.25deg_reg_v31.0e_sub2000.nc      # Subsampled precipitation mean (2000 time steps)
|   |   ├── rr_ens_spread_0.25deg_reg_v31.0e_sub2000.nc    # Subsampled precipitation spread (2000 time steps)
|   |   ├── rr_ens_mean_0.25deg_reg_v31.0e.nc              # Full precipitation mean dataset
|   |   ├── rr_ens_spread_0.25deg_reg_v31.0e.nc            # Full precipitation spread dataset
|   |   └── elev_ens_0.25deg_reg_v31.0e.nc                 # Elevation data
│   ├── data_utils.py                         # E-OBS data loading and processing
│   └── models.py                             # Self-supervised learning models
├── requirements.txt                          # Python dependencies
└── README.md                                 # This file

📊 Models and Architectures

TemporalPredictionModel

• Transformer-based with patch embeddings for spatial processing
• Multi-layer architecture with self-attention
• Temporal modeling for sequence prediction
• Lightweight design optimized for climate data

MaskedModelingModel

• U-Net architecture with encoder-decoder structure
• Skip connections for spatial detail preservation
• Flexible masking strategies for robust learning
• Batch normalization for stable training

🔧 Key Components

Data Loading (EOBSDataLoader)

• Loads E-OBS netCDF files with chunking for memory efficiency
• Handles precipitation mean and spread data
• Provides data information and statistics

Dataset Classes

• EOBSTemporalPredictionDataset: Creates temporal sequences for forecasting
• EOBSMaskedModelingDataset: Creates masked maps for reconstruction

Lightning Modules

• TemporalPredictionLightningModule: PyTorch Lightning wrapper for temporal models
• MaskedModelingLightningModule: PyTorch Lightning wrapper for masked models

🎓 Learning Outcomes

By exploring this repository, you will:

1. Understand self-supervised learning principles for climate data
2. Learn to work with large-scale precipitation datasets
3. Implement transformer-based temporal prediction models
4. Build U-Net architectures for spatial reconstruction
5. Use PyTorch Lightning for efficient training workflows
6. Evaluate model performance on real climate data

💡 Key Insights

Self-supervised learning with weather data presents unique challenges:

• Weather chaos: Small changes can lead to large differences
• Non-linear dynamics: Precipitation patterns are complex
• Multi-scale patterns: From local to synoptic scales
• Missing context: Need temperature, pressure, humidity for full picture

This repository demonstrates both the potential and limitations of self-supervised learning for climate data.

🔗 References

• E-OBS Dataset
• Copernicus Climate Change Service
• PyTorch Lightning

---

Self-supervised learning for climate science - bridging AI and atmospheric physics 🌍