Wildfires are among the most destructive natural disasters, causing severe damage to biodiversity, wildlife habitats, and air quality. This project utilizes Deep Learning and Computer Vision to develop an automated early detection system.

By analyzing aerial and ground-level forest images, the system predicts wildfire risks and detects early signs of fire/smoke, aiming to minimize false alarms and improve response times in protected forest environments.

• Multi-Model Analysis: A comparative study of 6 different neural network architectures.
• Robust Preprocessing: Implementation of ImageNet normalization, resizing to $224\times224$, and extensive data augmentation (flips, rotations).
• Bias Mitigation: Stratified sampling and balanced datasets to ensure ethical modeling and fairness.
• High Performance: The final model achieves a 99.41% F1-Score, prioritizing recall to ensure safety.

The model was trained on a combined dataset to ensure diversity and generalization:

1. Primary Dataset: 2,700 high-resolution RGB images (Wildfire Dataset). https://www.kaggle.com/datasets/elmadafri/the-wildfire-dataset
2. Supplemental Dataset: 5,050 additional images (Forest Fire Images) to reduce overfitting. https://www.kaggle.com/datasets/mohnishsaiprasad/forest-fire-images/data
3. Split: Data was divided into Training (70%), Validation (15%), and Test (15%) sets.

The project implemented and evaluated the following architectures:

Baseline Models:

• Multi-Layer Perceptron (MLP): A custom fully connected network built from scratch.
• AlexNet: A standard 8-layer CNN implemented in PyTorch.

Transfer Learning Models (Fine-Tuned):

• VGG16: Fine-tuned for binary classification.
• ResNet50: Utilized for capturing deep, complex patterns.
• MobileNetV2: Tested for efficiency in resource-constrained environments.
• InceptionV3: The best-performing model, utilizing multi-scale convolution modules.

The models were evaluated based on Accuracy, Precision, Recall, and F1-Score. The InceptionV3 (Fine-Tuned with SGD) was selected as the final model due to its superior balance of precision and recall.

Data derived from the Comparative Analysis Table.

• Language: Python
• Frameworks: PyTorch, NumPy
• Visualization: Matplotlib
• Environment: Jupyter Notebook / Google Colab

Perera B.P.N.

• Student ID: IT24100327
• Faculty of Computing, SLIIT UNI

This project is licensed under the MIT License.