Deep Learning on Aerial Imagery: What does it look
like on a map?
Rob Emanuele (@lossyrob)
What we’ll cover
• Describe the ISPRS Potsdam image
segmentation contest
• Brief overview of some Deep Learning
concepts
• View and analyze results of model
predictions on a map
GeoTrellis
• a Scala library for geospatial data types and
operations.
• enables Spark to work with geospatial data
for raster processing at scale
• provides lots of raster processing capability
for web servers and distributed computation
DSM generation
DEEP LEARNING
WHAT IS DEEP
LEARNING?
σ
x
1
x
3
y
Inputs (xi) are weighted
by the wis
The neuron “fires” if the
summed activation (x·w)
is above a threshold,
giving y
[In reality, y = σ(x·w),
where σ is some
“thresholding” function]
w1
w2
w3
x
2
Artificial Neurons
https://commons.wikimedia.org/wiki/File:Artificial_neural_network.svg
Artificial Neural Network (ANN)
Deep Neural Network (DNN)
…
Deep Neural Network (DNN)
…
THE DEEP PART
Convolutional Neural Network (CNN)
https://commons.wikimedia.org/wiki/File:Typical_cnn.png
Fully Convolutional Network (FCN)
Source: https://lmb.informatik.uni-freiburg.de/Publications/2015/RFB15a/
U-Net: Convolutional Networks for Biomedical
Image Segmentation
EXPERIMENT 1:
U-NET
EXPERIMENT 2:
FCN w/ IRRG
EXPERIMENT 3:
CROSS VALIDATION LATE
FUSION
ENSEMBLE
LET’S EXPLORE THE MAP
FUTURE WORK
Future Raster Vision Work
• Other CV tasks: Image Recognition, Object
Detection
• Infrastructure/Workflow improvements: utilize
AWS Batch for distributed training over
hyperparameters
• Integration into
• Inference on the fly, on the map or through an
API
THANK YOU
Rob Emanuele (@lossyrob)

Deep Learning on Aerial Imagery: What does it look like on a map?