Aerial Imagery dataset for fire detection: classification and segmentation using Unmanned Aerial Vehicle (UAV)
Title
FLAME (Fire Luminosity Airborne-based Machine learning Evaluation) Dataset
Paper
You can find the article related to this code here at Elsevier or
You can find the preprint from the Arxiv website.
Dataset
-
The dataset is uploaded on IEEE dataport. You can find the dataset here at IEEE Dataport or DOI. IEEE account is free, so you can create an account and access the dataset files without any payment or subscription.
-
This table below shows all available data for the dataset.
-
This project uses items 7, 8, 9, and 10 from the dataset. Items 7 and 8 are being used for the "Fire_vs_NoFire" image classification. Items 9 and 10 are for the fire segmentation.
-
If you clone this repository on your local drive, please download item 7 from the dataset and unzip in directory /frames/Training/... for the Training phase of the "Fire_vs_NoFire" image classification. The direcotry looks like this:
Repository/frames/Training
├── Fire/*.jpg
├── No_Fire/*.jpg- For testing your trained model, please use item 8 and unzip it in direcotry /frame/Test/... . The direcotry looks like this:
Repository/frames/Test
├── Fire/*.jpg
├── No_Fire/*.jpg- Items 9 and 10 should be unzipped in these directories frames/Segmentation/Data/Image/... and frames/Segmentation/Data/Masks/... accordingly. The direcotry looks like this:
Repository/frames/Segmentation/Data
├── Images/*.jpg
├── Masks/*.png- Please remove other README files from those directories and make sure that only images are there.
Model
- The binary fire classifcation model of this project is based on the Xception Network:
- The fire segmentation model of this project is based on the U-NET:
Sample
- A short sample video of the dataset is available on YouTube:
Requirements
- os
- re
- cv2
- copy
- tqdm
- scipy
- pickle
- numpy
- random
- itertools
- Keras 2.4.0
- scikit-image
- Tensorflow 2.3.0
- matplotlib.pyplot
Code
This code is run and tested on Python 3.6 on linux (Ubuntu 18.04) machine with no issues. There is a config.py file in this directoy which shows all the configuration parameters such as Mode, image target size, Epochs, batch size, train_validation ratio, etc. All dependency files are available in the root directory of this repository.
- To run the training phase for the "Fire_vs_NoFire" image classification, change the mode value to 'Training' in the config.py file. Like This
Mode = 'Training'
Make sure that you have copied and unzipped the data in correct direcotry.
- To run the test phase for the "Fire_vs_NoFire" image classification, change the mode value to 'Classification' in the config.py file. Change This
Mode = 'Classification'
Make sure that you have copied and unzipped the data in correct direcotry.
- To run the test phase for the Fire segmentation, change the mode value to 'Classification' in the config.py file. Change This
Mode = 'Segmentation'
Make sure that you have copied and unzipped the data in correct direcotry.
Then after setting your parameters, just run the main.py file.
python main.py
Results
- Fire classification accuracy:
- Fire classification Confusion Matrix:
- Fire segmentation metrics and evaluation:
- Comparison between generated masks and grount truth mask:
- Federated Learning sample
To consider future challenges, we defined a new sample of federated learning on a local node (NVidia Jetson Nano, 4GB RAM). Jetson Nano is available in two versions: 1) 4GB RAM developer kit, and 2) 2GB RAM developer kit. In this Implementation, the 4GB version is used with the technical specifications of a 128-core Maxwell GPU, a Quad-core ARM A57 @ 1.43 GHz CPU, 4GB LPDDR4 RAM, and a 32GB microSD storage. To test Jetson Nano for the federated learning, items (9) and (10) from Dataset are used for the fire segmentation. Since Jetson Nano has limited RAM, we assumed that each drone has access to a portion of the FLAME dataset. Only 500 fire images and masks are considered for the training and validation phase on the drone. As we aimed at learning a model on a smaller subset of the FLAME dataset and inferring that model, the default Tensorflow version is used here. Also, the image and mask dimension for each input is reduced to 128 x 128 x 3 rather than 512 x 512 x 3. To save more memory on the RAM, all peripherals were turned off and only WiFi was working at that time for the Secure Shell (SSH) connection. The setup of this node is:
Citation
If you find it useful, please cite our paper as follows:
@article{shamsoshoara2021aerial,
title={Aerial Imagery Pile burn detection using Deep Learning: the FLAME dataset},
author={Shamsoshoara, Alireza and Afghah, Fatemeh and Razi, Abolfazl and Zheng, Liming and Ful{\'e}, Peter Z and Blasch, Erik},
journal={Computer Networks},
pages={108001},
year={2021},
publisher={Elsevier}
}
Other related repositories and articles
License
For academtic and non-commercial usage