🌀 Typhoon Risk Analyzer — Western Pacific Ocean

Predicting Daily Typhoon Risk Using Oceanic & Atmospheric Data

An end-to-end machine learning system that predicts daily typhoon risk across the Western Pacific Ocean using NOAA OISST, ERA5, and IBTrACS datasets. The project integrates data preprocessing, feature engineering, XGBoost-based risk prediction, explainable AI (SHAP), and an interactive Streamlit dashboard.

🚀 Live Demo: https://typhoon-risk-analyser.streamlit.app/

📂 GitHub Repository: https://github.com/Bharanikumar55/Typhoon-Risk-Analysis

---

📌 Overview

Typhoons are among the most destructive natural disasters in the world, forming over warm ocean waters in the Western Pacific. This project builds an end-to-end data science pipeline — from raw scientific datasets to a deployed interactive web application — that predicts daily typhoon risk for any location in the Western Pacific region.

Unlike typical ML projects that use pre-cleaned Kaggle datasets, this project works with raw scientific reanalysis and satellite data used by actual climate researchers, requiring significant domain knowledge and data engineering effort.

---

🎯 Key Highlights

• Multi-source data fusion — integrates 3 independent scientific datasets (NOAA OISST, ERA5, IBTrACS) with different spatial resolutions and time frequencies
• Custom spatiotemporal label generation — typhoon presence labels generated using Haversine distance formula applied to storm track coordinates
• 10 years of data — trained on 2011–2019, tested on completely unseen year 2020 (genuine temporal generalization)
• Physically meaningful results — SHAP analysis confirms the model learned real meteorological patterns (low pressure = typhoon)
• Interactive deployment — Streamlit app allows exploration of any date from 2011–2020

---

🌍 Region of Study

Parameter	Value
Region	Western Pacific Ocean
Latitude	5°N – 45°N
Longitude	120°E – 160°E
Time Period	2011 – 2020 (June–November)
Spatial Resolution	1° × 1°
Temporal Resolution	Daily

This region covers the Philippine Sea, East China Sea, and waters surrounding Japan — the most active typhoon basin in the world.

---

📊 Data Sources

Dataset	Source	Variables	Resolution
NOAA OISST v2.1	https://www.ncei.noaa.gov/products/optimum-interpolation-sst	Sea Surface Temperature (SST)	0.25° × Daily
ERA5 Reanalysis (ECMWF)	https://cds.climate.copernicus.eu/	Mean Sea Level Pressure (MSL), 10 m U-Wind, 10 m V-Wind	0.25° × Hourly
IBTrACS v4	https://www.ncei.noaa.gov/products/international-best-track-archive	Tropical Cyclone Tracks	6-hourly

Study Period: 2011–2020 (June–November)

Study Region: Western Pacific (5°N–45°N, 120°E–160°E)

🧠 Machine Learning Pipeline

Raw Data (NetCDF)
      │
      ▼
┌─────────────────────────────────────┐
│  Phase 1: Data Preprocessing        │
│  • SST: crop region, filter months  │
│  • ERA5: Pa→hPa, compute wind speed │
│  • ERA5: hourly → daily averages    │
└─────────────────┬───────────────────┘
                  │
                  ▼
┌─────────────────────────────────────┐
│  Phase 2: Dataset Alignment         │
│  • Coarsen SST: 0.25° → 1°         │
│  • Interpolate ERA5 onto SST grid   │
│  • Merge into unified env dataset   │
└─────────────────┬───────────────────┘
                  │
                  ▼
┌─────────────────────────────────────┐
│  Phase 3: Label Generation          │
│  • Filter IBTrACS: WP basin         │
│  • Haversine distance per grid cell │
│  • Label = 1 if storm within 300km  │
└─────────────────┬───────────────────┘
                  │
                  ▼
┌─────────────────────────────────────┐
│  Phase 4: ML Training               │
│  • Logistic Regression              │
│  • Random Forest                    │
│  • XGBoost (best model)             │
│  • Temporal train/test split        │
└─────────────────┬───────────────────┘
                  │
                  ▼
┌─────────────────────────────────────┐
│  Phase 5: Deployment                │
│  • Streamlit interactive app        │
│  • Daily risk maps                  │
│  • Predicted vs actual comparison   │
└─────────────────────────────────────┘

---

📈 Results

Model Comparison (Single Year — 2011)

Model	AUC-ROC
Logistic Regression	0.9145
Random Forest	0.8924
XGBoost	0.9350 ✅

Final Model Performance (Multi-year — tested on 2020)

Metric	Value
AUC-ROC	0.86
Typhoon Recall	0.64
Training Period	2011–2019
Test Period	2020 (unseen)

The model was deliberately tested on a completely unseen year (2020) rather than a random split — this demonstrates genuine temporal generalization, a much stronger evaluation than standard cross-validation.

---

ROC Curve — All Models

All three models significantly outperform random guessing. XGBoost (green) achieves the best performance, hugging the top-left corner most closely.

---

Class Imbalance

Typhoon-present cells represent only ~2.4% of all data — a severe class imbalance. This was addressed using scale_pos_weight in XGBoost and class_weight='balanced' in other models, combined with stratified splitting.

---

Feature Importance

Both XGBoost and Random Forest consistently rank Sea Level Pressure as the dominant predictor (~65% importance in XGBoost). This aligns with meteorological theory — typhoons are extreme low-pressure systems.

---

SHAP Analysis

SHAP values confirm the model learned physically meaningful patterns:

• Low pressure (blue dots, left) → strong positive typhoon signal
• High wind speed (red dots, right) → positive typhoon signal
• High SST (red dots) → weak positive signal

The pressure dependence plot reveals a clear threshold around 1010 hPa — below this, typhoon risk increases sharply. This is consistent with known meteorology.

---

Typhoon Risk Maps

Daily predicted risk vs actual typhoon labels:

The model successfully identifies the circular low-pressure structure of a typhoon — a pattern it learned purely from data, without being explicitly programmed with any meteorological rules.

Average risk across the entire 2011 typhoon season:

The highest average risk zone (deep red) is centered around 15–30°N, 120–130°E — the Philippine Sea — which is the most active typhoon genesis region in the Western Pacific. The model correctly learned this spatial pattern from historical data.

---

🗂 Project Structure

typhoon_project/
│
├── data_raw/
│   ├── sst/                    # NOAA OISST NetCDF files (2011–2020)
│   ├── era5/                   # ERA5 reanalysis NetCDF files (2011–2020)
│   └── ibtracs/                # IBTrACS storm track data
│
├── data_processed/
│   ├── environmental_YYYY_jun_nov.nc   # Combined SST+MSL+Wind (per year)
│   ├── labels_YYYY_jun_nov.nc          # Typhoon labels (per year)
│   ├── X_multiyear.npy                 # Final ML feature matrix
│   ├── y_multiyear.npy                 # Final ML labels
│   └── xgb_final_model.pkl             # Trained XGBoost model
│
├── notebooks/
│   ├── era5.ipynb                  # ERA5 preprocessing
│   ├── sst.ipynb                   # SST preprocessing
│   ├── align_sst_era5.ipynb        # Dataset alignment
│   ├── labels_ibtracs.ipynb        # Label generation
│   ├── 05_dataset_construction.ipynb
│   ├── 06_baseline_models.ipynb
│   ├── 07_shap.ipynb
│   ├── 08_risk_map.ipynb
│   └── 09_multiyear_pipeline.ipynb
│
├── app.py                      # Streamlit deployment
├── requirements.txt
└── README.md

---

🚀 Run Locally

1. Clone the repository

git clone https://github.com/bharanikumar55/typhoon-risk-analyzer
cd typhoon-risk-analyzer

2. Install dependencies

pip install -r requirements.txt

3. Download data

Download the following datasets and place in data_raw/:

• NOAA OISST SST: PSL NOAA
• ERA5 Reanalysis: Copernicus CDS — variables: u10, v10, msl; region: 5°N–45°N, 120°E–160°E
• IBTrACS: NCEI IBTrACS

4. Run notebooks in order

era5.ipynb → sst.ipynb → align_sst_era5.ipynb → labels_ibtracs.ipynb
→ 05_dataset_construction.ipynb → 06_baseline_models.ipynb
→ 07_shap.ipynb → 08_risk_map.ipynb → 09_multiyear_pipeline.ipynb

5. Launch the app

streamlit run app.py

---

📦 Requirements

xarray
numpy
pandas
matplotlib
cartopy
scikit-learn
xgboost
shap
streamlit
joblib
netCDF4

---

🔬 What Makes This Project Unique

Aspect	This Project	Typical Student Project
Data source	Raw scientific NetCDF (ERA5, NOAA, IBTrACS)	Pre-cleaned Kaggle CSV
Label generation	Custom Haversine-based spatial labeling	Pre-made labels
Train/test split	By year (temporal)	Random shuffle
Evaluation	AUC-ROC, F1, temporal generalization	Accuracy only
Explainability	SHAP values with physical validation	None
Deployment	Live Streamlit app	None

---

👤 Author

Bharani Kumar Madisetty

---

📄 License

This project is licensed under the MIT License.

---

Built with ❤️ using ERA5 · NOAA OISST · IBTrACS · XGBoost · Streamlit