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

Sat4Survive - Saving lives using Deeplearning & Satellites

Miguel
January 18, 2019

Sat4Survive - Saving lives using Deeplearning & Satellites

There is a migrant crysis in the mediterranean, with hundreds of thousands persons moving to Europe looking for safety and a better live, every year and many drown in the sea.
Making aware everybody of this problem, is imperative to prevent it, this deck presents the proof of concept built using satellites ESA Copernicus public data coupled with artificial intelligence (Convolutional Neural Networks) in order to detect and alert every citizen when a small boat full of people is adrift in the mediterranean, particularly in Lybia.
- Presentation part of Saturdays.ai Demo Day in Barcelona on January 18th 2019.

Miguel

January 18, 2019
Tweet

More Decks by Miguel

Other Decks in Technology

Transcript

  1. Sat 4 Survive by SANTIAGO FRIAS MARIANNA POLINI JAN CARBONELL

    MIGUEL GUERRERO www.saturdays.ai @aisaturdayses https://github.com/sfrias/sat4survive Inspired by
  2. Some Variables - Just from Sat Sentinel-2: 1 TB per

    day, combination of the 290 km swath with 13 spectral channels incorporating 4 visible & near-infrared bands at 10m resolution per pixel - Infrared radiation of a human body is close to 100W, taking into account position, clothes and environmental conditions this translates into a reduction of about 80% of the IR radiation detectable per person - In our use case of detecting adrift small boats, we can expect IR radiation per pixel going from 3W to 50W
  3. Convolutional Neural Network (CNN) Resnet U-Net VGG trying* several CNN

    models * pro advice: go to Kaggle to find out + fast.ai updated library
  4. Example for military resolution: FRONTEX Public access resolution: Sentinel-2 even

    with significant less resolution, we still can get significant results ~lives
  5. ➢ All images trained in the model are 16 *

    16 pixels ➢ Resnet implemented is for bigger resolution ➢ Need to customize the number of K for a new resnet model ➢ Yes, it is feasible to implement a real-time early alert system (mini Humanitarian FRONTEX?) but still much work to do + resources ➢ More sat imagery sources (+radar, +bands) and improved “revisit” time Implementation
  6. The Dark Side ➢ We talked about saving lives with

    AI ➢ Also serves for “Search & Track” weaponry ➢ Plus a Deep learning (AI) model that “predicts”, “prescribes”, choose? targets …