🔬 Scientific Foundation
Research & Methodology
Exploring advanced deep learning techniques for robust forest fire and smoke detection. From dataset curation to "Learning without Forgetting".
Forest Fire C4 Dataset
A comprehensive, balanced dataset designed for environmental monitoring and fire detection tasks. Curated to handle complex scenarios like distinguishing smoke from cloud cover.
Comprehensive Data
A balanced collection of 4,823 images, meticulously labeled to ensure model fairness and accuracy across all conditions.
Total Images
4,823
Classes
4
Image Size
250x250
Format
JPEG
Fire
Visible flames and active fire fronts
No Fire
Clear forest, vegetation, and non-fire scenes
Smoke
Plumes and early-stage smoldering without flames
SmokeFire
Complex scenes containing both smoke and fire
Published Research
Learning without Forgetting
Published in Fire Ecology (Springer Open)
"Our core research focuses on "Learning without Forgetting" (LwF), a deep learning paradigm that enables models to adapt to new tasks—such as distinguishing between similar visual patterns like smoke and fog—without losing proficiency in previously learned categories. This approach is crucial for deploying robust AI in dynamic natural environments where conditions constantly evolve."
Read Full Paper
DOI: 10.1186/s42408-022-00165-0
Citation
Naren, O.S. Forest fire and smoke detection using deep learning-based learning without forgetting. Fire Ecol 19, 16 (2023).Model Architecture Difference
The model deployed in this live classifier is optimized for edge performance. Unlike the heavy research model described in the paper, this version uses a lightweight MobileNet architecture trained via Teachable Machine to ensure sub-second inference speeds via our API.
Technical Approach
The technical pipeline behind the results
1
Data Augmentation
Applied rigorous transformations including rotation, shearing, and brightness adjustments to ensure model robustness against lighting and orientation changes.
2
Transfer Learning
Leveraged pre-trained convolutional neural networks (CNNs) to extract high-level features, significantly reducing training time while improving accuracy.
3
LwF Implementation
Integrated Learning without Forgetting loss functions to preserve knowledge of "safe" forest states while aggressively learning new fire signatures.