This project aims to classify satellite imagery into Fire and Non-Fire categories using Transfer Learning (MobileNetV2). The goal is to support early detection of forest fires as part of broader efforts in ecosystem longevity, environmental sustainability, and disaster mitigation.
Forest fires cause severe ecological damage, including biodiversity loss, soil degradation, air pollution, and climate acceleration. By leveraging Deep Learning, this project attempts to automate early fire detection from satellite images, enabling faster response and reducing long-term ecological impacts.
This project uses:
- MobileNetV2 (ImageNet weights)
- Binary Classification
- Data Augmentation (Flip, Rotate, Zoom)
- Evaluation Metrics: Accuracy, Loss, Confusion Matrix, F1-score
The dataset is sourced from Kaggle:
- Forest Fire Dataset
- Contains two classes:
FireandNon-Fire - Images resized to
224 x 224
Note: Dataset is not included in this repo due to size. Download it manually from Kaggle and place it in:
/data/raw
/docs # laporan, roadmap, evaluasi
/notebooks # .ipynb (Colab notebook)
/results # output evaluasi (cmatrix, plots)
/models # final model (.h5 / .tflite)
git clone https://github.com/Deluxe8841/.git cd
pip install -r requirements.txt
| Metric | Value (Estimation) |
|---|---|
| Accuracy | 85–92% |
| F1-Score | High (balanced) |
accuracy_loss.png
confusion_matrix.png
- Use YOLO or Faster R-CNN for detection (not just classification)
- Use Sentinel / MODIS time-series
- Convert model to TFLite for mobile deployment
- Add geospatial visualization (GIS)
If you use this project, cite:
@article{mobilenetv2,
title={MobileNetV2: Inverted Residuals and Linear Bottlenecks},
author={Sandler, Mark et al.},
booktitle={CVPR},
year={2018}
}
Dataset:
- Kaggle Forest Fire dataset