Wildfire Detection using Remote Sensing & Machine Learning

Problem Statement

Wildfires cause significant environmental and economic damage worldwide.
Accurate and fast detection of burned areas is critical for disaster response and ecological assessment.

Traditional methods are time-consuming and not scalable.
This project aims to automate burned area detection using satellite imagery and machine learning.

Project Description

This project focuses on detecting wildfire-affected areas using Sentinel-2 satellite imagery and machine learning models in Google Earth Engine.

The system analyzes post-fire satellite data to classify burned areas and evaluate model performance.

---

Case Study: Manavgat Wildfire (2021)

This project demonstrates the methodology using the Manavgat wildfire event in Turkey.
However, the pipeline is designed to be adaptable to different geographic regions and wildfire events.

---

Technologies Used

• Google Earth Engine (GEE)
• JavaScript
• Sentinel-2 Satellite Data
• Machine Learning:
  • Random Forest
  • Support Vector Machine (SVM)

---

How It Works

1. Satellite imagery is filtered (low cloud coverage)
2. Training data is labeled:
   • Burned areas
   • Healthy vegetation
   • Water (for improved classification)
3. Models are trained and evaluated
4. Burned areas are classified and visualized
5. Total burned area is calculated (hectares)

---

Model Comparison (Random Forest vs SVM)

Results:

• Random Forest Accuracy: 0.9757

• SVM Accuracy: 0.9079

• RF Burned Area: 35,794 ha

• SVM Burned Area: 53,060 ha

• Difference: 17,265 ha

Key Insight:

Random Forest produces more reliable and stable results, while SVM tends to overestimate burned areas.

---

Error Analysis (Impact of Water Class)

Results:

• SVM V1 Accuracy: 0.9925
• SVM V2 Accuracy: 0.9967

Confusion Matrix:

V1 (Without Water Class): [[1063,19], [20,4140]]

V2 (With Water Class): [[992,47,0], [65,4182,0], [0,0,29650]]

Key Insight:

Including water as a separate class significantly reduces misclassification and improves model robustness.

---

Visual Results

dNBR & Burn Severity

Burned Area Detection

Post-Fire Satellite Image

---

Model Comparison

---

Error Analysis

---

Project Structure

images/ dnbr-severity-map.png burned-area-detection.png post-fire-image.png model-rf.png model-svm.png model-comparison.png svm-v1.png svm-v2.png error-difference.png

src/ dNBR-analysis.js data-preprocessing.js random-forest-classification.js model-comparison.js error-analysis.js

---

How to Run

1. Open Google Earth Engine Code Editor
2. Copy the scripts from the src/ folder
3. Run the code
4. Visualize results on the map

---

Future Improvements

• Integrate Deep Learning models (CNN)
• Automate wildfire detection pipeline
• Add real-time monitoring capability
• Build a web-based visualization dashboard

---

Project Highlights

• Remote sensing + machine learning integration
• Model comparison (RF vs SVM)
• Error analysis with class improvement
• Real-world wildfire case study
• Scalable to different geographic regions

---

Contact

If you’d like to collaborate or discuss this project, feel free to reach out!