🌧️ Cloudburst Risk Prediction System — Uttarakhand

A production-grade machine-learning system that predicts day-ahead cloudburst risk for 20 Himalayan hotspots across all 13 districts of Uttarakhand, India. Trained on 36 years of NASA POWER meteorological data and validated against 6 famous historical disasters including Kedarnath 2013.

🔗 Live Demo: https://cloudburst-uttarakhand-b2xeqpso8fhxzs8kkjpp2n.streamlit.app/
📦 Repository: https://github.com/UTK-BAJPAI/cloudburst-uttarakhand

---

🎯 What This Project Does

• 🧠 Honest predictive modelling — no target leakage, next-day forecast horizon
• 🌍 20 hotspots, all 13 districts of Uttarakhand
• 🇮🇳 Bilingual UI — English / हिंदी toggle
• 🗺️ Interactive map with real-time risk markers
• 🕐 Historical backtest against 6 famous Uttarakhand events
• ⚡ Live data from Open-Meteo (free, no API key)
• ☁️ Cloud deployed — 24/7 public access on Streamlit Cloud

---

📊 Historical Validation

The model was tested on famous past Uttarakhand cloudburst events. Below is the honest, reproducible validation:

#	Event	Date	NASA Rain	Model	Outcome
1	Kedarnath Disaster	16 Jun 2013	116 mm	85.8% High	✅ Caught
2	Mandakini Valley Flash Flood	17 Jun 2013	57.1 mm	71.4% High	✅ Caught
3	Pauri Cloudburst	14 Aug 2012	130 mm	68.6% High	✅ Caught
4	Mussoorie Cloudburst	12 Aug 2009	0.2 mm*	67.6% High	✅ Caught (via lag features)
5	Joshimath Glacier Disaster	7 Feb 2021	1.9 mm	15.1% Low	✅ Correctly flagged low (non-cloudburst)
6	Tehri Cloudburst	13 Aug 2003	14.8 mm†	56.8% Medium	⚠ Borderline (grid-dilution limitation)

* Grid-averaged value; localized point rainfall was much higher
† NASA POWER's 0.5° × 0.625° grid dilutes localized cloudbursts

Aggregate metrics:

• Recall on positives: 4 / 5 = 80%
• Specificity on negatives: 1 / 1 = 100%
• 5-fold CV ROC-AUC: 0.71 ± 0.05

---

🏗️ Architecture

flowchart LR
    A[NASA POWER<br/>1990-2026] --> B[data_pipeline.py<br/>20 sites × 13K days]
    B --> C[Balance + Label<br/>cloudburst_dataset.csv]
    C --> D[Feature Engineering<br/>15 lag/rolling features]
    D --> E[Random Forest<br/>5-fold CV, AUC 0.71]
    E --> F[model.pkl + scaler.pkl]
    F --> G[Streamlit GUI]
    H[Open-Meteo<br/>Live API] --> G
    G --> I[Public URL<br/>24/7 access]
    F --> J[Power BI<br/>Dashboard]

Loading

---

🚀 Quick Start

# Clone
git clone https://github.com/UTK-BAJPAI/cloudburst-uttarakhand.git
cd cloudburst-uttarakhand

# Set up Python 3.12 venv
python -m venv .venv
.\.venv\Scripts\Activate.ps1     # Windows
# source .venv/bin/activate      # Linux/Mac

# Install
pip install -r requirements.txt

# Build dataset (5 min from NASA POWER API)
python data_pipeline.py --use-api

# Train model
python train_model.py

# Generate static figures
python visualizations.py

# Launch interactive app
python -m streamlit run app.py

Then open http://localhost:8501 in your browser.

---

🛠️ Tech Stack

Layer	Tools
ML / Data	scikit-learn (RandomForest), pandas, numpy
Data Sources	NASA POWER (historical), Open-Meteo (live)
UI	Streamlit, PyDeck (interactive map)
BI	Power BI Desktop with custom dark theme
Hosting	Streamlit Community Cloud (free), GitHub
Automation	Windows Task Scheduler (hourly live updates)

---

🔬 Methodology

Honest Predictive Modelling (No Target Leakage)

• Target: Cloudburst on day t+1
• Features: 15 features from days t and earlier — including lag-1 and 3-day rolling means
• Threshold: Fixed at 0.5 for balanced production metrics
• Validation: Stratified 5-fold cross-validation

Synthetic Injection Discovery

During development, I discovered that the original pipeline injected synthetic cloudburst-like features (rainfall ≥ 110mm, humidity ≥ 88%) on labelled event dates if real measurements were below threshold. This was inflating model accuracy artificially.

Fix: Disabled synthetic injection, kept only naturally-occurring extreme rainfall as positive labels. Joshimath 2021 (a glacier collapse, not a cloudburst) was removed from event labels.

Result: Honest 5-fold AUC of 0.71 (vs inflated 0.94 before fix). Joshimath now correctly predicts 15.1% Low risk instead of 91% High.

---

⚠️ Honest Limitations

1. NASA POWER grid resolution. The data uses a 0.5° × 0.625° grid (~55 × 65 km), which spatially averages localized cloudbursts. Tehri 2003 dropped 200+mm at the disaster point but only 14.8mm grid-averaged → model gave borderline 56.8%.
2. Class imbalance. Cloudbursts are rare; balanced dataset is 506 rows after down-sampling.
3. Day-ahead horizon only. No multi-day forecast yet.

---

🛣️ Future Work

• Ensemble with IMD AWS data for sub-grid resolution
• 3-day forecast mode using Open-Meteo's forecast endpoint
• Telegram bot alerts for HIGH risk sites
• Time-series CV (train pre-2015, test post-2015) for true out-of-sample validation
• Probability calibration with Platt scaling
• Mobile-first UI for field officers in affected districts

---

📁 Project Structure

---

👤 Author

Utkarsh Bajpai
📧 askus.utkarsh@gmail.com
🔗 GitHub: UTK-BAJPAI

---

📜 Citation

If you use this work, please cite:

@misc{bajpai2026cloudburst,
  author       = {Bajpai, Utkarsh},
  title        = {Cloudburst Risk Prediction System for Uttarakhand},
  year         = {2026},
  howpublished = {\url{https://github.com/UTK-BAJPAI/cloudburst-uttarakhand}},
  note         = {Live demo at cloudburst-uttarakhand.streamlit.app}
}

---

🙏 Acknowledgments

• NASA POWER for free historical meteorological data (1990-2026)
• Open-Meteo for free real-time weather API (no key required)
• Streamlit Community Cloud for free hosting
• IMD for cloudburst event records used in validation

---

📜 License

MIT — see LICENSE.

---

Built with ❤️ for the Himalayas. 13 districts, 20 sites, 36 years of data — one honest model.