Deep Dive into Deep Learning
Rohit Singh
Vinay Viswambharan
Vinay Viswambharan Rohit Singh
Who are we?
Session Overview
• Deep Learning in ArcGIS
• Types of Deep Learning Models
- Imagery
- Video
- 3D Point Clouds
- Feature layers and tabular data
- Unstructured Text
• The Road Ahead
• Resources
Deep Learning in ArcGIS
Different data types, tasks and integration
Pixel
Classification Object
Point Detection
Text Oriented
Cloud
Imagery Super
Resolution Object
Classification
Feature
Attachments Data Types Tasks
Motion
Imagery
Satellite Prediction &
Imagery Classification
LiDAR
Instance
Segmentation
Bathymetric Entity Extraction
Aerial Point Cloud
Data Classification
Imagery
Integration
Why is deep learning important?
Traditional human data interpretation doesn’t scale….
More sensors
Large volumes of imagery
Velocity of data
Automation
Accuracy
Detecting
Shipwrecks
Deep Learning in ArcGIS Pro
ArcGIS ♥ Notebooks
Text goes here
ArcGIS API for Python
arcgis.learn module
The arcgis.learn module in ArcGIS API for Python enables Python
developers and data scientists to easily train and use deep learning
models with a simple, intuitive API.
Train
Models
Before After
• Dozens of lines of • 3-5 lines
Code • Easy Installation
• Installing External DL (ArcGIS Pro &
Frameworks Notebooks)
• HARD! • EASY
Land Cover
Classification
Deep Learning with arcgis.learn
Deep Learning Models in ArcGIS
Object Classification Assign a label to a given image / geographical feature
Model:
- FeatureClassifier
Applications: Damaged Undamaged
- Damaged building classification
- Clean or ‘green’ pools…
- Broken/flashed insulators
- Identifying plant species
Tree Inventory
Field data collection with Survey123
Pixel Classification Assign a label to each pixel
Models:
- UNetClassifier
- PSPNetClassifier
- DeepLab
Applications:
- Land Cover Classification
- Pervious/Impervious mapping
- Building Footprint Extraction
- Cloud detection
- Road detection
- Water Meters from oblique images
Land Cover
Classification
Using sparse training data
Object Detection Find objects and their locations (bounding boxes)
Models:
- SingleShotDetector
- RetinaNet
- YOLO
- FasterRCNN
Applications:
- Detect trees, cars, airplanes, swimming pools, well
pads
- Detect encroaching structures
- Moving objects like road cracks/catfish on video
- Road signs, and other assets in oriented imagery
- Detecting and counting vehicles
- Shipwrecks in Bathymetric data
- Fire hydrants in Oriented imagery
Faster R-CNN
Gold Standard in Object Detection
YOLO:
You Only Look Once
Ready to use model for Oriented Imagery
Pavement Crack
Detection
Applying object detection to video
Object Detection Find objects and their precise locations (masks or polygonal features)
Models:
- MaskRCNN
Applications:
- Building footprint extraction
- 3D roof reconstruction
- Sinkholes
3D Building
Reconstruction
Deep Learning on LiDAR data
Point Cloud Classification
Models:
- PointCNN
Applications:
- Classifying point cloud data
(buildings/ground, etc)
- Power line and utility poles extraction
Point Cloud
Classification
3D Deep Learning with Point Clouds
Image-to-Image Translation
Models:
- SuperResolution
Applications:
- Improved image interpretability
- Improved image resolution
SuperResolution
Improve quality and resolution of
imagery
Feature and tabular data
Models:
- Fully Connected Network
- MLModel
Applications:
- Dengue prediction
- Solar power plan efficiency
prediction
- Integrating ArcGIS with Machine
Learning models from scikit-learn
Unstructured Text – Natural Language Processing
Models:
- EntityRecognizer
Applications:
- Parsing unstructured text and convert to
spatial information
- Mapping crime from police reports
- Extracting fire incidents from fire reports
Crime incident
extraction
Extracting and analyzing crime
incidents from unstructured text
Geospatial AI: The Road Ahead
Ready-to-use Models
Building Footprints Extraction
Cloud Removal
Cloud removal using Image Inpainting
Key Takeaways
ArcGIS has powerful deep learning capabilities!
• Suite of tools for Deep Learning workflows
• All imagery categories/tasks and more..
• Variety of training model types
• Variety of clients
• Massively scalable
• Powerful API for developers and data scientists
• Robust image management to complement deep learning
• Rich tools to perform down stream analysis
Resources
Section Subhead
Additional resources
arcgis.learn module in ArcGIS API for Python Deep learning in ArcGIS Pro
► go.esri.com/uc2020/deep-learning-help
► developers.arcgis.com/python
► Samples
► Guide YouTube Videos
► API Reference ► Search for “ArcGIS Deep Learning”
GitHub Repo
► github.com/arcgis-python-api
Medium GeoAI Blog
► medium.com/geoai