ML-Enhanced Fog Forecasting using WRF Post-Processing

This repository accompanies the research manuscript:

"Weather Research and Forecasting (WRF) model and machine learning algorithms to improve marine fog predictions at two coastal locations in Atlantic Canada"

The project demonstrates how lightweight machine learning (ML) models can be used as a post-processing layer on Numerical Weather Prediction (NWP) output to significantly improve hourly fog forecasts out to 24 hours, without modifying physical parameterizations within the Weather Research and Forecasting (WRF) model.

---

📌 Research Overview

Marine fog remains difficult to predict using traditional NWP due to challenges in representing microphysics, boundary-layer turbulence, and air–sea interactions. WRF often overpredicts surface liquid water content, leading to persistent false fog events.

This study reframes fog prediction as a binary classification problem and applies ML models to correct systematic WRF biases at station scale. The approach was evaluated at two fog-prone coastal sites in Atlantic Canada:

• St John’s, Newfoundland and Labrador
• Yarmouth, Nova Scotia

Using 12 years of reanalysis-based training data and independent pseudo-operational WRF forecasts for summer 2024, ML post-processing consistently outperformed a WRF-only fog diagnostic.

---

🧠 Key Contributions

• Hybrid physics-informed ML post-processing framework
• 24-hour hourly fog forecasts, exceeding typical ML nowcasting horizons
• Comparison of tree-based ensembles and deep learning approaches
• Robust statistical evaluation including McNemar’s test and bootstrap confidence intervals
• Operationally interpretable results suitable for decision support

---

🧪 Machine Learning Framework

Fog occurrence is defined using visibility observations (≤ 1 km) and modeled as an imbalanced binary classification problem.

Models Evaluated

• ExtraTrees Classifier (ETClassifier) - primary operational model
• XGBoost Classifier
• Bidirectional LSTM (biLSTM)

Tree-based ensemble methods were found to be more robust and interpretable than deep learning, with comparable or better performance.

---

🌍 Data Sources

Training Data (2012–2023)

• ERA5 reanalysis (hourly)

  • 2 m temperature (T2)
  • 2 m relative humidity (RH2)
  • 10 m wind components (U10, V10)
  • Surface pressure (P_sfc)

• NAV CANADA visibility observations

Forecast Data (2024)

• WRF v4.5.2 pseudo-operational simulations
• Initialized using GFS 0.25° forecasts
• 36 h daily runs with 12 h spin-up

• Preprocessed ERA5 data to downscale to 9 km to match the WRF simulations for more accurate comparisons.
• 2-nested domain used for WRF Simulations (do1=27 km, do2=9km).
• Used NCL to extract the features used from the ERA5 dataset and WRF (get_variable.ncl).

---

🧩 Feature Engineering

Features used by ML models include:

• Meteorological variables at forecast time
• 1-hour lagged predictors to capture persistence
• Time-based features (hour, month)

Feature importance analysis shows relative humidity and its persistence dominate fog predictability, with wind direction playing a secondary role.

---

📊 Evaluation Metrics

Model performance is assessed using:

• Precision
• Recall
• F1-score
• Confusion matrices
• McNemar’s statistical test
• Bootstrap-derived 95% confidence intervals

ML post-processing improved F1-score by:

• +13% at St John’s
• +18% at Yarmouth

relative to WRF liquid-water-based fog detection.

• Confusion matrices comparing ETClassifier and WRF-only fog detection at St John’s and Yarmouth (2024).
• ETClassifier reduces false alarms and increases fog detection skill compared to WRF-only diagnostics at both sites.

---

📈 Key Results

• ETClassifier achieved the most balanced skill across sites
• ML models reduced both false alarms and missed fog events
• Forecast skill remained stable across the full 24-hour horizon
• Statistically significant improvement over WRF (p < 0.05)

---

🧠 ML Perspective

This work positions ML as a bias-correction tool, not a replacement for physical modeling. Results show that:

• Feature selection matters more than model complexity
• Ensemble tree methods outperform deep learning when inputs are physically meaningful
• ML post-processing is computationally efficient and operationally deployable

---

📁 Repository Structure

fog-ml-wrf/
├── era5 data/Binary    # era5 data collected for both locations
├── 2024_test/    # 2024 forecasted features for both locations
├── WRF_Files/    # configurations used for preprocessing ERA5 data and running WRF + extracting the variables
├── ML_scripts/
│   ├── fogtest_decisiontree.py    # preprocessing and training models (ETClassifier) + evaluating metrics
│   └── fogtest_neural.py    # Using neural nets as another option
├── Paper_figures/    # contains all the plots used in manuscript
└── README.md

---

⚙️ Environment

• Python ≥ 3.9
• scikit-learn
• xgboost
• tensorflow / keras
• numpy, pandas, matplotlib
• Linux-based HPC to run WRF
• NCL scripts

---

🚀 Operational Relevance

The proposed workflow is well-suited for:

• Aviation and marine decision support
• Hourly fog alerts with extended lead time
• Integration into existing NWP-based forecasting systems

---

📖 Citation

If you use this work, please cite:

Teeloku, P., Chen, Z., Taylor, P., & Chen, Y. (2025).
Weather Research and Forecasting (WRF) model and machine learning algorithms to improve marine fog predictions at two coastal locations in Atlantic Canada.
Under review.

---

📬 Contact

For questions or collaboration inquiries:

Author: Piyush Teeloku

Email: piyush31@yorku.ca

---

📝 License

This project is intended for academic and research use.