Sensing the World

1. Mike Warren
Co-founder & CTO
Descartes Labs
@m_warren
Sensing the World

2. A dynamic model of the whole Earth
allows us to
Find
Find things, measure
them, and monitor
change over time.
Predict
Collect – Integrate – Model
Predict future changes
with increasing
confidence.
Act
Act to minimize risk and
optimize outcomes, with
feedback.

3. Predicting Corn Yields

4. Predicting supply and demand in
dynamic supply chains
Fields Storage Bins Grain Elevators Processors Consumers

5. Geospatial scientists spend 80% of their time preparing data;
Descartes Labs eliminates the dominant source of friction.
– Data analysts spend 80% of their time simply finding and preparing data.
– Many satellite images are difficult to access; not "online"; must be requested
and downloaded.
– One satellite constellation may not be adequate; e.g., resolution, frequency,
sensor band.
– Archives contain raw, unprocessed images; lots of clouds, not registered or
calibrated.
– Machine learning requires clean and reproducible datasets, often in
combination with supercomputing capability.

6. Our geospatial analysis platform
A growing archive of analysis-ready images,
with historical records for back-testing
models.
Robust pipeline for continuous updating as
new images become available.
Multiple satellite image datasets integrated
into a single system.
Cloud-based data and services, ready-to-use
for data science teams anywhere.
Elastic computing model—scaling compute
resources quickly and only when needed.

7. Descartes maps
Landsat 8, RGB Sentinel-2, Red EdgeSentinel-1, SAR
15 meters per pixel
3.1 trillion pixels per band
Created from 70 trillion pixels per
band collected from 2013 to 2017
(320 TB)
20 meters per pixel
680 billion pixels
Created from Synthetic Aperture
Radar (SAR) range/azimuth
measurements collected from
2014 to 2017 (86 TB)
20 meters per pixel
1.8 trillion pixels per band
Created from 22 trillion pixels per
band collected from 2015 to 2016
(120 TB)

10. GeoVisual search
t-SNE visualization of image chips
Divide the earth's surface
into (billions of) small,
overlapping images
Extract a "visual feature
vector" from each image
using a convolutional
neural network
Search for neighbors in this
feature space using either
direct search or locality-
sensitive hashing

12. https://www.descarteslabs.com/beta-application.html
If you’d like to learn more,
“Data-Intensive Supercomputing in the Cloud: Global Analytics for Satellite Imagery”
arXiv:1702.03935 [cs.DC]
https://medium.com/descartestech
https://www.descarteslabs.com
Join our Platform Beta
Thank you!