Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial...
Transcript of Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial...
![Page 1: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/1.jpg)
Imagery, text and geospatial Machine Learning
applications in Montreal's booming ML
landscape
Tom Landry1, Samuel Foucher1,
Mario Beaulieu1, Jean-François Rajotte1
(1) CRIM
ESGF Face to Face 2017
San Francisco, 2017-12-07
![Page 2: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/2.jpg)
![Page 3: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/3.jpg)
![Page 4: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/4.jpg)
![Page 5: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/5.jpg)
![Page 6: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/6.jpg)
![Page 7: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/7.jpg)
Non-for-profit organisation circa 1985
8.2M$ budget
ISO 9001:2008 certified
4 R&D teams:
- Vision and imaging
- Emerging technologies and data
science
- Speech and text
- Advanced software modelling and
development
2016-2017 by the numbers
- 95 projects
- 166 clients
- 50 academic collaborators
- 11 scientifc seminars
- 41 scientific publications
- 50 employees
![Page 8: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/8.jpg)
● Creation of time series from the 64 x 3 files
● Featuring (mean, std...)
● Principle Component Analysis (Spark ML)
● K-means (Spark ML)
● Daily 10km gridded dataset
● Precipitation & Temperature (1950-2013)
Scalable Machine Learning Using SciSpark
![Page 9: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/9.jpg)
![Page 10: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/10.jpg)
It is a mixture of density and grid-based clustering algorithm. It has linear complexity and near linear
horizontal scalability. As a result, PatchWork can cluster a billion points in a few minutes only, a 40x
improvement over Spark MLLib native implementation of the well-known K-Means
Highly Scalable Grid-Density Clustering
Algorithm for Spark MLLib
![Page 11: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/11.jpg)
11Vision Géomatique 2017
• ‘Deep Features’ approach:
• Uses an already trained network to produce features
• A classifier is added
• Differerent CNN trained with machine vision data *CaffeNet, GoogleNet,
etc.)
Classification and detection
![Page 12: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/12.jpg)
Results on Pleiades imagery (50 cm)
![Page 13: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/13.jpg)
Vision Géomatique 2017
▪ ConvNet used to estimate missing high frequencies
Ref: J. Kim, J. K. Lee and K. M. Lee, "Accurate Image Super-Resolution Using Very Deep Convolutional Networks," 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, 2016, pp. 1646-1654.
Super resolution techniques
![Page 14: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/14.jpg)
Vision Géomatique 2017
5 to 2.5 meters per pixel
10 to 2.5 meters per pixelLow resolution (bicubic) Super-resolution High resolution
Gain of about 10% in performance
![Page 15: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/15.jpg)
Other applications of ML at CRIM
- Real-time speech transcription
- Vocal biometry
- Firewall logs access event detection
- Anormal event detection in airports
- People and car tracking
- Fire services response time
- Bike rental sources and destination
- Personalized product recommandation
- Active learning for user queries
- Many, many more...
![Page 16: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/16.jpg)
Canada’s plan for 2018-2021
Cyberinfrastructure challenge. Results known next March.
![Page 17: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/17.jpg)
![Page 18: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/18.jpg)
![Page 19: Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel](https://reader030.fdocuments.us/reader030/viewer/2022041015/5ec63bbed176c0630e62c3ba/html5/thumbnails/19.jpg)