Scribd
Upload
0 views

01_Seminar_Introduction

The document introduces BlackSky, a geospatial intelligence company focused on satellite imagery and machine learning analytics. It outlines the seminar's agenda, covering topics such as satellite orbits, image processing, and deep learning techniques applicable to satellite imagery. Additionally, it discusses the applications of machine learning in analyzing satellite data for various tasks like change detection and object detection.

Uploaded by

melek.sahlia
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0 views

01_Seminar_Introduction

The document introduces BlackSky, a geospatial intelligence company focused on satellite imagery and machine learning analytics. It outlines the seminar's agenda, covering topics such as satellite orbits, image processing, and deep learning techniques applicable to satellite imagery. Additionally, it discusses the applications of machine learning in analyzing satellite data for various tasks like change detection and object detection.

Uploaded by

melek.sahlia
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 29

Deep Learning with Satellite Imagery

Seminar Introduction
BlackSky Introduction
● Geospatial intelligence company with offices in Herndon, VA and Seattle, WA.
● Building constellation of small sats with plans of orbiting over 20 satellites.
● Mid-inclination orbits yield exceptionally high revisit rates over targets.
● Imagery derived analytics are core part of our mission.
● Growing data science team with heavy specialization in machine learning.

BlackSky Orbits
BlackSky Introduction
With machine learning, predictive
algorithms, and natural language
processing, BlackSky delivers critical
geospatial insights about an area or topic
of interest. We synthesize data from a
wide array of sources including social
media, news outlets and even
earthquake sensors.
BlackSky Introduction

Patrick O’Neil Diego Torrejon

● Director of Machine Learning and Artificial ● Data Scientist at BlackSky.


Intelligence at BlackSky. ● Working at BlackSky for three years.
● Working at BlackSky for four years. ● PhD in mathematics from GMU.
● PhD in mathematics from GMU. ● Former researcher at NIST.
● Worked at GeoEye prior to BlackSky. ● Recipient of the NSF Graduate Research
● Background in topological data analysis and Fellowship.
probabilistic modeling.
Seminar Overview
1. Seminar Overview
2. Satellite Imagery (Orbits, Bands, Collection, Resolution, etc)
3. Satellite Image Processing (Orthorectification, Color Correction, Dynamic Range
Adjustment, Co-registration, etc.)
4. Classical Computer Vision (Convolutions, Sobel Filters, Watershed and Segmentation)
5. Deep Learning Overview (Deep Neural Networks, CNNs, RNNs, etc)
6. Deep Learning Ingredients (Dropout, Batch Normalization, Vanishing Gradient, etc.)
7. Convolutional Neural Networks
8. U-Net and W-Net
9. ResNet and ODEs
10. Generative Adversarial Neural Networks.
Satellite Imagery
Getting to Orbit
● Rocket Labs Electron
○ 150–225 kg to 550 SSO
○ $6M per launch
● ISRO PSLV
○ 1,750 kg to SSO
○ $21M-31M per launch
● SpaceX Falcon 9
○ 22,800 kg to LEO
○ $50M per launch
● ULA Atlas V
○ 20,520 kg to LEO
○ $109M per launch
● CNES Ariane 5
○ Over 20,000 kg to LEO
○ $165M-220M per launch
Orbital Terminology
● Remote sensing requires an understanding of
where a satellite is in its orbit.
● Apoapsis and Periapsis: The furthest and closest
points respectively, of an orbit with respect to its
host.
● Apogee and Perigee: Same as above with the host
being Earth.
Orbital Mechanics Terminology
● Inclination describes the angle of an
orbital plane with respect to a reference
plane.
● For Earth the reference plane is taken to
be the equatorial plane.
Orbital Mechanics Terminology
● Orbital eccentricity describes the amount by
which an orbit deviates from a perfect circle.
● All satellites we will be discussing have an
eccentricity below one and usually pretty close
to zero.
Earth 0.0167

Mars 0.0934

Jupiter 0.0484

Halley's Comet 0.9671

ʻOumuamua 1.20
Orbital Terminology
Semi-major Axis:

Orbital Velocity:

Eccentricity:

Delta-V for Inclined Orbit: v2 dv


𝞡
v1
Important Types of Orbits
● Low-Earth Orbit (LEO): Lower than 2000 Kilometers
○ Optimal for remote sensing due to proximity to Earth
○ Small field of view
○ Sun-Synchronous Orbit
● Geosynchronous Earth Orbit (GEO)
○ Orbital period matches Earth’s rotation on its axis
○ GEO is at a 35,786 km altitude above the Earth
○ Useful for communications satellites
● Orbits around Earth are often described by their
Two-Line Element Set (TLE).
BLACKSKY GLOBAL 2
1 43812U 18099BG 19040.20317408 .00000356 00000-0 37510-4 0 9998
2 43812 97.7571 113.6950 0014563 47.2297 313.0146 14.95340708 9661

BlackSky TLE
Satellite Imagery: Major Players
● Globals (BlackSky)
● Worldview Constellation (DigitalGlobe)
● Sentinel (ESA)
● LandSat (NASA & USGS)
● Pleiades (Airbus)
● Kompsat (SIIS)
● Dove & SkySat (Planet)
Characteristics of Satellite Imagery
● Spectral Bands: Satellite imagery may contain
additional bands beyond the visible spectrum
(e.g. SWIR, NIR, Thermal IR, etc).
○ In a tensor framework, multiple band imagery is
viewed as third order tensors.
● Ground Sample Distance: the distance
between pixel centers* measured on the
ground (related to resolution).
● Observation Angle. The angle at which the
image was taken relative to the ground.
Applications of Remote Sensing
● Mapping missions.
● Prediction of crop yield (right).
● Military surveillance.
● Rescue missions.
● Size estimation of oil spills.
● Monitoring ice caps.
● Natural disaster assessment.
● Fresh water estimation (bottom).

Left: True soybean yield. Middle: Previous


Model. Right: Deep learning Model.
Geospatial Data Tools
● Raster vs Vector
● QGIS and Esri ArcGIS
● GDAL
● Rasterio
● geopy
Machine Learning
What is Machine Learning?
● The goal of machine learning
is to teach a computer
program to perform a task.
● Classification: To label an
email as spam or not.
● Regression: To estimate US
population in 2020.
● Dimensionality Reduction: To
extract keywords of a news
article.
● Clustering: To find cliques in
Facebook.
Types of Machine Learning
There are many types of machine learning, the following classification highlights the type of data utilized.

Supervised Unsupervised Semi-supervised


Training data has associated Training data doesn’t have Some training data has
labels. associated labels. associated labels.
Pro: Faster training and simpler
We want to learn the function We may attempt to learn gathering of data.
p(y | x) or p(x, y) p(x) Con: Models tend to be more
complex.
Example: Support vector Example: K-nearest neighbors Example: GAN + Classifier
machines (classification). (clustering).
What is Deep Learning?
● A Neural Network consist of multiple nodes which mimic the biological neurons
of a human brain.
● Deep learning uses a hierarchy of these layers, which serve to extract more
abstract representations/features from the data.
Why Machine Learning?

● Machine learning can now automate many tasks that were thought to be only
performed by humans (e.g. image recognition).
● Address the issue of needing a huge number of analysts to look at immense
quantities of data.
● Models have been trained to beat humans in board games such as chess and
go; and in live video games such as Dota 2 and Starcraft II.
● These accomplishments have shown that the AI models pick up on hidden
patterns in the data, previously missed by humans.
Machine Learning and Satellite Imagery
Machine learning can be applied to
satellite imagery in the following tasks:

● Change detection at a site of


interest.
● Object detection (buildings, ships,
planes, etc).
● Image segmentation (cities, roads,
water, forest, etc).
● Resolution enhancement of
imagery.
Before Image

Change Detection
Change Detection Image

After Image

Red: Objects in the before image and


not in the after image.
Blue: Objects in the after image and not
in the before image.
Object Detection

Mask R-CNN Architecture


Semantic Segmentation

U-Net Architecture
Imagery Enhancement

Original Image Enhanced Image


Python & Keras
Keras is a Python library which provides a layer
of abstraction for Tensorflow, PyTorch, and
other deep learning frameworks.

Examples in this seminar will use Keras.

NVIDIA GPUs can be used to drastically


increase the training and inference speed.
Example Keras Model
# For a single-input model with 10 classes (categorical classification):
model = Sequential()
model.add(Dense( 32, activation= 'relu', input_dim= 100))
model.add(Dense( 10, activation= 'softmax' ))
model.compile(optimizer= 'rmsprop' ,
loss= 'categorical_crossentropy' ,
metrics=[ 'accuracy' ])

# Generate dummy data


import numpy as np
data = np.random.random(( 1000, 100))
labels = np.random.randint( 10, size=( 1000, 1))

# Convert labels to categorical one-hot encoding


one_hot_labels = keras.utils.to_categorical(labels, num_classes= 10)

# Train the model, iterating on the data in batches of 32 samples


model.fit(data, one_hot_labels, epochs= 10, batch_size= 32)

You might also like