Structural Damage Classification in Wildfire Events

A data-driven project aimed at predicting the extent of structural damage in wildfire events based on various building and environmental factors. The project utilizes machine learning techniques for classification and provides a user-friendly web interface for predictions.

---

---

Steps Followed in the Project

1. Problem Understanding

• Objective: Predict structural damage severity during wildfire incidents using building characteristics and geographical attributes.
• Type of Problem: Classification problem (predicting categorical damage levels).

---

2. Dataset Overview

The dataset includes the following features:

• CAL FIRE Unit: The region under CAL FIRE jurisdiction.
• Structure Type: Type of structure (e.g., single-family residence, commercial, etc.).
• Roof Construction: Material used for the roof (e.g., asphalt, metal, tile, etc.).
• Vent Screen: Mesh size of vent screens (e.g., <= 1/8", none, etc.).
• Exterior Siding: Type of exterior siding material (e.g., wood, stucco, etc.).
• Window Pane: Type of window glazing (e.g., single-pane, double-pane, etc.).
• Incident Start Date: Date of the wildfire event.
• Latitude & Longitude: Geographic coordinates of the structure.
• Damage: Target variable indicating the extent of structural damage.

Key Observations:

• Certain materials and construction types exhibit higher vulnerability to wildfires.
• Location-based factors may influence damage severity.
• Feature encoding was performed for categorical variables.

---

3. Data Preparation

1. Imported Libraries: Pandas, NumPy, Scikit-learn, Matplotlib, Seaborn.
2. Dataset Loading: Data was read, cleaned, and structured for modeling.
3. Missing Values: Checked and handled missing values appropriately.
4. Exploratory Data Analysis (EDA):
   • Visualized feature distributions and their relationship with damage levels.
   • Identified potential correlations and risk factors.

---

4. Feature Engineering

1. Handeled Incident Start Date: As the Machine Learning couldn't understand date foramt so created new columns seperately for Incident Start Year, Month & Day.
2. Categorical Encoding: Encoded categorical variables such as CAL FIRE Unit, Structure Type, Roof Construction, Vent Screen, Exterior Siding, Window Pane & Damage.
3. Scaling: Standardized numerical features (Latitude, Longitude).
4. Data Splitting: Split the dataset into training and testing sets for model evaluation.

---

5. Model Building

Implemented the following classification models:

1. Logistic Regression
2. Decision Tree Classifier
3. Random Forest Classifier
4. Gradient Boosting Classifier
5. XGBoost Classifier

---

6. Model Evaluation

• Evaluation Metrics:
  • Accuracy Score
  • Precision, Recall, and F1-Score
  • Confusion Matrix
• The models were compared, and the best-performing model was selected based on classification accuracy and recall for severe damage cases.

---

7. Results Visualization

1. Confusion Matrix: To understand misclassifications.
2. Feature Importance: Identified key predictors of structural damage.

---

8. Deployment

The best-performing model decision tree classifier model was saved and deployed as a web application:

• Model Saved:

  import joblib
  joblib.dump(best_model, 'wildfire_damage_model.pkl')

• Web Application:
  • Backend developed using FastAPI.
  • Frontend implemented with HTML, CSS, and JavaScript for user input and predictions.

---

Project Structure

project/
├── api/
│   ├── main.py                                   # FastAPI backend for handling predictions
├── data/
│   ├── california-wildfire-dataset.csv           # Dataset used for model training
├── frontend/
│   ├── index.html                                # Web interface for the application
│   ├── styles.css                                # Styling for the web page
│   ├── script.js                                 # Form handling and API interaction
├── model/
│   ├── Structural Damage Classification.ipynb    # Notebook for EDA and model training
│   ├── decision_tree_model.pkl                   # Saved machine learning model
    ├── encoded_data_columns.pkl
    ├── label_encoder.pkl
    ├── model_scaler.pkl                 
    ├── one_hot_encoders.pkl
    ├── ordinal_encoder.pkl
    ├── scaler.pkl
├── README.md                                     # Project documentation

---

How to Run the Project

Backend

1. Navigate to the api/ directory.
2. Run the main.py file using Python.
3. The backend will start at http://localhost:8000.

Frontend

1. Open the index.html file in the frontend/ directory.
2. Fill out the form with structure details and submit to get the predicted damage classification.

---

Future Scope

• Improve model accuracy by incorporating additional environmental variables.
• Implement GIS-based analysis for better risk assessment.
• Deploy the application to a cloud service for real-time wildfire damage predictions.

---