computer vision for satellite data applications · ground data availability ground data alignment...
TRANSCRIPT
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
Elecnor Deimos is a trademark which encompasses Elecnor Group companies that deal with Technology and Information Systems: Deimos Space S.L.U., Deimos Castilla La Mancha S.L.U., Deimos Engenharia S.A., Deimos Space UK Ltd., Deimos Space S.R.L. (Romania).
Computer Vision for Satellite Data Applications
Paula Martí
1
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd. 2
• Elecnor Deimos is the technology company of the Elecnor Group, operating in Aerospace, Defence, Transport, Energy, Environment, etc.
• Located in 4 countries, 300+ engineers:
o Spain (HQ), Portugal, UK, Romania
• Space is our core business. Experts in satellite software and systems engineering:
o Flight Systems
o Ground Systems & Space Situational Awareness
o Satellite Navigation
o Earth Observation & Satellite Data Applications
o Space Missions (Deimos-1,2)
Deimos Space UK
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
About me
33
Paula Martí
• Technical Manager at Deimos UK
o Remote sensing applications
o Satellite data processing
o Management and business development
• Education:
o MSc Telecommunications EngineeringUniversidad Politécnica de Valencia2007
o MSc Computer VisionOxford Brookes University, UK2015
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
Elecnor Deimos is a trademark which encompasses Elecnor Group companies that deal with Technology and Information Systems: Deimos Space S.L.U., Deimos Castilla La Mancha S.L.U., Deimos Engenharia S.A., Deimos Space UK Ltd., Deimos Space S.R.L. (Romania).
Why is computer vision important for the geoinformation world?
4
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
What we do with satellite data
Feature extraction and counting:
• Trees to help calculate insurance value
• Polar bears to ensure sustainable population levels
• Number of plants in a field to estimate yield
• Counting Graves to map cemeteries
• 3D modelling
5
Source: http://gisgeography.com/100-earth-remote-sensing-applications-uses/
Classifying the land:
• Extent of the different types of land and its use
• Mapping mineral deposits with hyperspectral data
Finding boundaries:
• Riparian areas: border between land and a river or a lake
• Forest fires
• Urban growth
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
Elecnor Deimos is a trademark which encompasses Elecnor Group companies that deal with Technology and Information Systems: Deimos Space S.L.U., Deimos Castilla La Mancha S.L.U., Deimos Engenharia S.A., Deimos Space UK Ltd., Deimos Space S.R.L. (Romania).
Satellite data applications using classic methods of Machine Learning
6
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
CropID: Crop Classification for the whole of UK
Satellite data plus geometry of the field
boundaries
Apply machine learning algorithms
(Random forest, SVM)
Classify by choosingthe most probable crop for each field
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
SAFIY: Urban Mapping
8
Objective:
Automatically update maps in Dubai using high-resolution satellite imagery
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
SAFIY: Road detection with SVM
9
Machine learning
algorithms to
classify road
pixels
Vectorise and
update mapChallenges
- Shadows are often
misclassified
- It might not work on
all types of roads
- Simple, straight
vectors are difficult
to extract
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
SAFIY: Feature extraction
Roundabout detection
1 – Create a feature descriptor that is able to ‘describe’ a roundabout
2 – Find circles in the image
3 – Classify the circles to discard round features that are not roundabouts
4 - Vectorise
10
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
Elecnor Deimos is a trademark which encompasses Elecnor Group companies that deal with Technology and Information Systems: Deimos Space S.L.U., Deimos Castilla La Mancha S.L.U., Deimos Engenharia S.A., Deimos Space UK Ltd., Deimos Space S.R.L. (Romania).
Modern computer vision techniques applied to satellite data
11
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
Urban 3D Reconstruction
12
Techniques:
• General-purpose 3D reconstruction from images: stereo, wide-baseline matching, camera calibration, and point triangulation
• 3D reconstruction using LIDAR
• Extrusion from footprint
• Computer Vision techniques such as Gaussian Process Latent Variable Models (GPLVM)
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
Urban 3D Reconstruction
13
GPLVM:
• Trained with 3D
models of the
buildings.
• Create a latent
variable space of what
a building looks like.
• Given the outline, the
best shape for the
building can be fitted.
• Only 1 view needed.
Input Input
Output Output
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
Road detection using Deep Learning
Apply Deep Learning algorithms to previous data
Training data:
We only used 5 Deimos-2 images cut into 48x48 pixel tiles (~100 tiles)
14
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
Road detection using Deep Learning – some results
15
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
Challenges
Ground data availability
Ground data alignment with satellite data
Deep learning is demanding in terms of processing power and memory
• 8-layer Fully Convolutional Network VGG provided by Oxford University
• Rented AWS most powerful machine with GPU and TensorFlow
• Train the network with small tiles
• Result: Machine runs out of memory!
16
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
Current projects
Palm tree detection Car counting
DMU-CMS-SUPSC03-PRE-11-E © DEIMOS Space UK Ltd.
Collaborations
We are happy to discuss:
- Detection of features of interest to other users
- Ground data
- Ways to easily label data
- Processing capabilities
- Experts in TensorFlow
- Different algorithms that you might want to try with our data and labels to compare results
18
20