Monitoring and identification of disasters are crucial for mitigating their effects on the
environment and on human population, and can be facilitated by the use of unmanned aerial vehicles
(UAV), equipped with camera sensors which can produce frequent aerial photos of the areas of interest. A
modern, promising technique for recognition of events based on aerial photos is deep learning. In this paper,
we present the state of the art work related to the use of deep learning techniques for disaster monitoring
and identification. Moreover, we demonstrate the potential of this technique in identifying disasters
automatically, with high accuracy, by means of a relatively simple deep learning model. Based on a small
dataset of 544 images (containing images of disasters such as fires, earthquakes, collapsed buildings,
tsunami and flooding, as well as “non-disaster” scenes), our preliminary results show an accuracy of 91%
achieved, indicating that deep learning, combined with UAV equipped with camera sensors, have the
potential to predict disasters with high accuracy in the near future. Presented at the EnviroInfo 2017 Conference in Luxembourg.
Scaling API-first – The story of a global engineering organization
Disaster Monitoring using Unmanned Aerial Vehicles and Deep Learning
1. 1
EnviroInfo Conference 2017
Disaster Management for Resilience
and Public Safety Workshop
Disaster Monitoring using UAV and Deep Learning
Andreas Kamilaris
13th September, 2017
Luxembourg
3. Motivation
3
Disaster monitoring can be facilitated by the use of
unmanned aerial vehicles (UAV), equipped with
camera sensors which can produce frequent aerial
photos of the areas of interest.
4. Motivation
4
Advantages of Drones:
• Small size
• Low cost of operation
• Exposure to dangerous environments
• High probability of mission success
• No risk of loss of aircrew resource
• High-resolution image sensing
• High operational flexibility
6. Motivation
6
Advantages of Deep learning:
• Superior performance in terms of precision
• Perform classification and predictions particularly
well due to their structure.
• Flexible and adaptable
• No need for hand-engineered features
• Generalizes well
• Robust in low-resolution and -quality images.
Andreas Kamilaris and Francesc X. Prenafeta-Boldú, Deep Learning in Agriculture: A
Survey, Computers and Electronics in Agriculture Journal, 2017. [Under review]
7. Research Questions
7
Can drones and aerial image sensing be used for
real-time monitoring of physical areas and?
accurate identification of disasters?
Can deep learning be used in combination with
drones and aerial images for real-time disaster
monitoring/identification?
9. Deep Learning
9
Convolutional Neural Networks
• Can be applied to any form of data, such as audio,
video, images, speech, and natural language.
• Various “successful” popular architectures: AlexNet,
VGG, GoogleNet, Inception-ResNet etc.
• Pre-trained weights
• Common datasets for pre-training CNN architectures
include ImageNet and PASCAL VOC.
• Many tools and platforms that allow researchers to
experiment with deep learning e.g. Keras, Theano.
11. State of the Art
11
No. Disaster Image source Accuracy
1.
Fire (Kim, Lee, Park, Lee, &
Lee, 2016)
Aerial photos
Human-like
judgement
2.
Avalanche (Bejiga, Zeggada,
Nouffidj, & Melgani, 2017)
Aerial photos 72-97%
3.
Car accidents and fire (Kang
& Choo, 2016)
CCTV cameras 96-99%
4. Landslides (Liu & Wu, 2016)
Optical remote
sensing
96%
5.
Landslides and flood (Amit,
Shiraishi, Inoshita, & Aoki,
2016)
Optical remote
sensing
80-90%
12. Methodology
12
CNN Model: VGG architecture, pre-trained with the
ImageNet dataset of images.
Dataset: 544 aerial photos from Google images (min.
256x256 pixels), acquired using the query:
[Disaster]: earthquake, hurricanes, flood and fire.
[Landscape]: aerial views of cities, villages, forests and
rivers
[Disaster | Landscape] + "aerial view" + "drone"
13. Dataset
13
No. Image Group
No. of
Images
Relevant Possible
Disaster
1. Buildings collapsed 101
Earthquakes and
hurricanes
2. Flames or smoke 111 Fire
3. Flood 125
Earthquakes,
hurricanes and
tsunami
4. Forests and rivers 104 No Disaster
5. Cities and urban landscapes 103 No Disaster
16. Setup
16
• 82% (444 images) of our dataset as training data
and 18% (100 images) as testing data.
• Random assignment of images in training/testing.
• Training procedure 20 minutes on a Linux
machine, testing 5 minutes for the 100 images.
• Learning rate: 0.001
• Used data augmentation techniques.
• 30 epochs
17. Results: Training Vs. Testing
17
83
84
85
86
87
88
89
90
91
92
82-18 70-30 75-25 85-15 90-10
Training Vs. Testing Percentage
OverallPrecision(%)
18. Results: Training Vs. Precision
18
0
10
20
30
40
50
60
70
80
90
100
5 10 15 20 25 30 35
OverallPrecision(%)
Number of Epochs
20. Results: Analysis of Error
20
9% Error
Urban Vs. Buildings collapsed (4%) Urban Vs. Fire (2%)
Urban Vs. Flooding (1%)Flooding Vs. Buildings collapsed (2%)
21. Conclusion
21
Deep learning offers good precision and many benefits.
Can be successfully used in combination with UAV for
disaster monitoring/identification.
It has also some disadvantages:
• It takes (sometimes much) longer time to train.
• It requires the preparation and pre-labeling of a
dataset containing at least some hundreds of images.
22. Future Work
22
• Publish the dataset to the research community.
• Enhance the dataset with more images.
• Experiment with different architectures, platforms and
parameters.
• Increase overall precision to more than 95%.
• Perform a real-life case study with drones used for
monitoring some particular disaster e.g. indication of
fire.
23. Vision
23
Better disaster modelling,
especially when combining UAV
and deep learning with geo-
tagging of the events identified
and geospatial applications.
Facilitate the integration of relevant actors (i.e. action
forces/authorities, citizens/volunteers, other stakeholders)
in disaster management activities with regard to
communication, coordination and collaboration.
24. 24
Many thanks for your attention!
Andreas Kamilaris
andreas.kamilaris@irta.cat
Notas del editor
If wireless objects (e.g. sensor networks) represent a future of “billions of information devices embedded in the physical world,” why should they not run the standard internetworking protocol?
Thus, the Internet can penetrate into the real world of physical objects.
If wireless objects (e.g. sensor networks) represent a future of “billions of information devices embedded in the physical world,” why should they not run the standard internetworking protocol?
Thus, the Internet can penetrate into the real world of physical objects.