🌲🔥 WiDS Global Datathon 2026: Wildfire Survival Prediction

Predicting immediate wildfire threats using early-stage kinematic data.

📌 Project Overview

The Challenge: When a wildfire ignites, emergency managers have a "golden window" to make life-saving decisions. Using data restricted to the first 5 hours of a fire, this project predicts the probability that a fire will reach a populated area (evacuation zone) within four critical time horizons: 12h, 24h, 48h, and 72h.

The Constraint: The dataset is extremely small (221 training events) and heavily right-censored (many fires never hit). This required a rigorous "Small Data" strategy rather than standard Deep Learning approaches.

🏆 The Strategy: "The Committee of Experts"

Instead of a single complex model, we deployed a Multi-Horizon Expert Framework. We treated each time horizon (12h, 24h, 48h, 72h) as a distinct survival problem, training specialized models for each.

1. Feature Engineering (The Physics Engine)

Raw data was insufficient. We engineered kinematic features to teach the model the laws of motion:

• est_time_to_contact: (Distance to Town) / (Closing Speed).
• growth_intensity: (Area Growth) / (Initial Area).
• threat_momentum: (Speed × Acceleration).

2. The Model Tournament

We rigorously compared two approaches using Stratified 5-Fold Cross-Validation:

• 🔴 Approach A (Generalist): One Regressor predicting exact "Time to Hit".
• 🟢 Approach B (Specialist): Four Classifiers (Random Forest), each answering "Will it hit in < X hours?"

The Result: The Specialist approach won by a landslide (AUC 0.95+ vs. 0.70), proving that simpler, focused models outperform complex ones on small, chaotic data.

3. Logical Post-Processing

We enforced Monotonicity constraints on the final predictions.

• Logic: A fire cannot be more likely to hit in 12 hours than in 24 hours.
• Algorithm: Prob(T < 24h) = max(Prob(T < 24h), Prob(T < 12h))

---

📊 Key Results

Our validation (Stratified K-Fold) yielded exceptional stability across all horizons:

Horizon	Model Type	AUC Score	Interpretation
12 Hours	Random Forest (Class)	0.958	Highly accurate at detecting immediate threats.
24 Hours	Random Forest (Class)	0.989	Near-perfect separation of "Safe" vs. "Danger".
48 Hours	Random Forest (Class)	0.996	Excellent long-term risk assessment.
72 Hours	Random Forest (Class)	1.000*	Note: Due to high censorship, all long-surviving fires in training eventually hit.

---

🛠️ Repository Structure

├── data/
│   ├── train.csv           # Training data (Features + Targets)
│   ├── test.csv            # Test data (Features only)
│   └── metaData.csv        # Column dictionary
├── notebooks/
│   └── wildfire_analysis.ipynb  # Full analysis: EDA, Validation, & Modeling
├── submission_final.csv    # FINAL OUTPUT for leaderboard
└── README.md               # This file

🚀 Quick Start

1. Prerequisites

• Python 3.8+
• pandas, numpy, scikit-learn, matplotlib

2. Running the Analysis

The entire pipeline is contained in the main notebook. It executes the following steps automatically:

1. Audit: Checks for censorship ratios and zero-variance columns.
2. Split: Sets up Stratified K-Fold (5 splits) to prevent data leakage.
3. Engineer: Calculates physics-based features.
4. Train: Trains 4 separate Random Forest Experts.
5. Predict & Fix: Generates probabilities and enforces monotonicity.

⚠️ Critical Note on Data

• Censorship: ~70% of the training fires never hit the town.
• Sample Size: Only ~50 fires hit within the first 12 hours.
• Action: We used class_weight='balanced' in all models to prevent the AI from ignoring the rare "Hit" events.