Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Helping humanitarian mapping through machine learning

Helping humanitarian mapping through machine learning

This talk has been delivered on FOSS4GFR 18.

Videos from this presentation:
- Denver https://youtu.be/BEvsUrk1N28
- Karthoum https://youtu.be/3K_-y0hCUJQ
- Ghana https://youtu.be/qjCFuEY5GmQ
- WebApp https://youtu.be/FCOsRhPFr_k

In order to provide assistance in humanitarian crisis, search & rescue teams needs up-to-date and accurate maps of affected areas. Many organizations assist them for this reactive need. Some are governmental (NATO, ESA..), others coordinate teams of volunteers in updating OpenStreetMap (CartONG, Map Give, Humanitarian OpenStreetMap Team, Missing Maps ...).

In such a reactive production, machine learning may be of great help to mappers. We have explored use cases of buildings detection and automatic extraction of building contours. We have tried to solve these problems with deep learning by combining OSM labelling and Pleiades & Spot imagery (CNES/Airbus DS).

In this talk, we will review our results, provide an overview on our methodology and its limitations, and share some insights on how to obtain better models.

Magellium

May 16, 2018
Tweet

More Decks by Magellium

Other Decks in Science

Transcript

  1. PUTTING KNOWLEDGE ON THE MAP Helping humanitarian mapping through machine

    learning Clément Maliet Sébastien Bosch @seeb0h
  2. Who are we? Clément Maliet engineer signal and image processing,

    spatial and aerial imagery, ML, coding, R&D...
  3. Who are we? Sébastien Bosch - @seeb0h engineer geomatics/maps, spatial

    and aerial imagery, robotics, ML, coding, R&D, mgmt...
  4. Humanitarian mappping ? Basey, Samar after the devastation wrought by

    Typhoon Haiyan (local name: Yolanda) on November 8, 2013. Lawrence Ruiz CC-BY-SA-4.0 Humanitarian Crisis Typhoon
  5. Humanitarian Crisis Earthquake Images of Port-au-Prince, Haiti, Jan 15, 2010.

    Department of Defense assets have been dispatched to Haiti to assist with humanitarian assistance and disaster relief after a magnitude 7 earthquake hit the country on Jan. 12, 2010 Tech. Sgt. James L. Harper Jr., USAF Public domain
  6. Humanitarian Crisis War Images of Port-au-Prince, Haiti, Jan 15, 2010.

    Department of Defense assets have been dispatched to Haiti to assist with humanitarian assistance and disaster relief after a magnitude 7 earthquake hit the country on Jan. 12, 2010 Tech. Sgt. James L. Harper Jr., USAF Public domain
  7. Humanitarian mappping ? Humanitarian Crisis Population Displacement The Sahrawi refugees

    – a forgotten crisis in the Algerian desert EU Civil Protection and Humanitarian Aid Operation CC BY-ND 2.0
  8. Many challenges for humanitarian responders. Among them : • Where

    are the people – Gathered together ? Isolated ? – Which density ?
  9. Many challenges for humanitarian responders. Among them : • Where

    are the people – Gathered together ? Isolated ? – Which density ? • How to reach them – Where are the transportation infrastructure? – What their condition is ?
  10. Collaborative mapping Haïti 10/01/10 From Crowdsourced Mapping to Community Mapping:

    The Post-Earthquake Work of OpenStreetMap Haiti Robert Soden and Leysia Pale
  11. Collaborative mapping Haïti 10/02/05 From Crowdsourced Mapping to Community Mapping:

    The Post-Earthquake Work of OpenStreetMap Haiti Robert Soden and Leysia Pale
  12. Preemptive mapping “To map the most vulnerable places in the

    developing world, in order that international and local NGOs, and individuals can use the maps and data to better respond to crises affecting the areas. “
  13. Machine Learning Learning to reproduce a result, a behavior Icons

    made by http://www.freepik.com from www.flaticon.com is licensed by CC 3.0 BY
  14. Machine Learning Learning to reproduce a result, a behavior Traditional

    Programming data program output Icons made by http://www.freepik.com from www.flaticon.com is licensed by CC 3.0 BY
  15. Machine Learning Learning to reproduce a result, a behavior Traditional

    Programming data program output data program output Machine Learning Icons made by http://www.freepik.com from www.flaticon.com is licensed by CC 3.0 BY
  16. program estimates evaluation model fitting loop Icons made by http://www.freepik.com

    from www.flaticon.com is licensed by CC 3.0 BY data output Machine Learning Learning to reproduce a result, a behavior
  17. Machine learning to make footprints Isolate patches around buildings Generate

    buildings footprint Machine Learning task: From mapping to ML
  18. Machine learning to make footprints Isolate patches around buildings Generate

    buildings footprint Machine Learning task: From mapping to ML
  19. Machine learning to make footprints Isolate patches around buildings Generate

    buildings footprint Machine Learning task: Detection From mapping to ML
  20. Machine learning to make footprints Isolate patches around buildings Generate

    buildings footprint Machine Learning task: Detection Segmentation From mapping to ML
  21. But… • Better models for temperate climate • Not enough

    data to generalize • Use 30cm imagery
  22. Our needs : • A generalized model • Better-suited for

    sub-tropical area • Use OneAtlas imagery
  23. Footprint extraction World-wise model OneAtlas setting Vector mask annotation Diverse

    training set Dataset specifications OneAtlas Pleiades target Context
  24. Vector mask annotation Diverse training set Dataset specifications OneAtlas Pleiades

    target Datasets Image source Satellite Native resolution Working resolution SpaceNet challenge Worldview 3 31 cm 59 cm NOAA public Haiti image Worldview 3 31 cm DSTL challenge Worldview 3 31 cm BD ORTHO® 50 cm Pleiades 50 cm OneAtlas + OSM mask Pleiades 59 cm Trainset Image source Satellite Native resolution OneAtlas Pleiades 59 cm Testset
  25. It is significantly easier to manually delete false positives rather

    than to visually check the image to add back false negatives.
  26. Training/Optimization Not a challenge winner: • General model • High

    recall • Visual test inspection to alleviate annotation noise
  27. Training/Optimization Not a challenge winner: • General model • High

    recall • Visual test inspection to alleviate annotation noise Training:
  28. Training/Optimization Not a challenge winner: • General model • High

    recall • Visual test inspection to alleviate annotation noise Training: • Early stopping
  29. Training/Optimization Not a challenge winner: • General model • High

    recall • Visual test inspection to alleviate annotation noise Training: • Early stopping • Stochastic Gradient Descent
  30. We constructed a general model able to segment buildings in

    most sub-tropical regions of the world
  31. We constructed a general model able to segment buildings in

    most sub-tropical regions of the world It shows consistent performances in various contexts on a diverse set of building appearances.
  32. We constructed a general model able to segment buildings in

    most sub-tropical regions of the world It shows consistent performances in various contexts on a diverse set of building appearances. However…
  33. Some technical challenges are yet to be addressed: • Output

    mask quality and precision • Instance separation
  34. Some technical challenges are yet to be addressed: • Output

    mask quality and precision • Instance separation • Vector mask generation from instance raster mask
  35. Some technical challenges are yet to be addressed: • Output

    mask quality and precision • Instance separation • Vector mask generation from instance raster mask • OSM annotation quality control
  36. Some technical challenges are yet to be addressed: • Output

    mask quality and precision • Instance separation • Vector mask generation from instance raster mask • OSM annotation quality control • Vector registration (global and local) • Vector quality assessment and correction
  37. Industrial pipeline integration is yet to be addressed: • How

    to integrate efficiently and seamlessly in a mapper’s workflow
  38. Industrial pipeline integration is yet to be addressed: • How

    to integrate efficiently and seamlessly in a mapper’s workflow • How the model will behave in a real mapping situation
  39. Industrial pipeline integration is yet to be addressed: • How

    to integrate efficiently and seamlessly in a mapper’s workflow • How the model will behave in a real mapping situation • Refine necessary strengths and point out weaknesses
  40. Industrial pipeline integration is yet to be addressed: • How

    to integrate efficiently and seamlessly in a mapper’s workflow • How the model will behave in a real mapping situation • Refine necessary strengths and point out weaknesses • Iterate the process with active learning