Aerial Forest Fire Detection using Deep Learning

Problem Statement

Early detection of forest fires is critical for containment. Satellite latency is often too high, necessitating low-altitude aerial surveillance (drones) processed by efficient edge-compatible models.

Methodology

We designed a custom Convolutional Neural Network (CNN) architecture optimized for binary classification (Fire / No-Fire).

Data Preprocessing

To reduce false alarms (often caused by sunsets or red foliage), we implemented aggressive augmentation:

Color Jittering: Varying saturation to force the model to learn texture, not just color.
Random Rotation: Ensuring orientation invariance.

Model Architecture

model = Sequential([
    Conv2D(32, (3, 3), activation='relu', input_shape=(224, 224, 3)),
    MaxPooling2D(2, 2),
    
    Conv2D(64, (3, 3), activation='relu'),
    MaxPooling2D(2, 2),
    
    Flatten(),
    Dense(128, activation='relu'),
    Dropout(0.5), # Prevent overfitting
    Dense(1, activation='sigmoid')
])

Performance Metrics

The model was evaluated on a dataset of 10,000+ aerial images.

Metric	Value
Accuracy	> 90%
False Alarm Reduction	30%
Inference Time	45ms

Conclusion

By optimizing the batch training process and refining the dataset, we created a model viable for deployment on drone hardware with limited GPU overhead.