JPL Machine Learning For Earth Science
JPL Machine Learning For Earth Science Dr. Lukas Mandrake (JPL 398, 8x/NSTA) CL#: 17-0845 Copyright 2017 California Institute of Technology. U.S. Government sponsorship acknowledged.
2
Data Science Life Cycle Decision Autonomy Support
Archiving
Knowledge
Analysis
o Quick Analysis • Quick Calibration • Quick QC • Quick Analysis • Quick Knowledge o Annotate o Advise / React o Repeat
QC
Curation
Calibration
Operations
Transport
Acquisition
Formulation
Autonomy / Advisory Loop
Data Mining
Quick-Look products enable earlier reaction Advisory systems direct focus
3
What is Machine Learning? Algorithms that inductively self-assemble from examples. Expert Examples
Functional System
( e.g. Estimates 1-10 of attractiveness of picture )
• • • •
Training Data ( Encompasses domain of responsibility )
Interpretable as required Fixed code (V&V ready) Not ever-learning (mostly) EASILY UPDATED
Strength: Don’t specify rules Don’t explain “how” Machine Learning simplifies & systematizes the building and 4 updating of Autonomy / Advisory systems
Higher Level Questions / Actions “Where should I start looking?” “Show me more like this.” “What is likely to happen next?” “(Re)optimize my system.” ”How many kinds are there?” “Show me the most interesting first.” “What inputs are most informative?” “Show me new things.” 5
1. Science-driven use cases 2. Explanability 3. “Let Me Help” 6
Drs. Lukas Mandrake, Gary Doran, Brian Bue
Extremophiles
DHM field team Nuuk, Greenland
Raw Hologram (2D) Detections Contamination
• Digital Holographic Microscopes • Big data (4D, ~GB/s), rare findings • Motility ~ Life (composition agnostic!) • HELM ML system detects, tracks, and classifies in messy, raw 2D holograms
Global Surveys
7
Dr. David Thompson, Dr. Brian Bue, et al
CH4 detection in four corners
Enabled ground team to find underground pipe leaks
Airborne Imaging Spectrometer Multiple gas pipelines shut down / repaired Machine Learning “That Matters”
8
Dr. Lukas Mandrake Don’t Pre-Filter Data: No Data Quality Flags
instead
Order by Trust Learn from Experts
Current OCO-2 Quality Estimation Product / NASA Software of the Year runner-up
9
SMACC Endmembers Superpixel segmentation
Similar orbital tech used for • • • •
Minerology maps Identifying crop-types Recognizing diseased citrus Estimating hurricane damage
D. Thompson et al., 2012. TGARS.
10
• Borup Fiord sulfur springs • Biosignature analog site for Europa • Detect and track from orbit (EO-1) • •
Support vector machine (SVM) classifier 26 detections as of May 31, 2016 Courtesy: Steven E. Grasby
EO-1 image (Hyperion)
Sulfur detection (yellow) D. Gleeson et al., 2010. Remote Sensing of Environment. 11 L. Mandrake et al., 2012. ACM TIST.
12
Seedling Concepts • $30-$50k • Proof of Principle • Single Problem
Demo System • $150k-$300k • Extensive Validation • Single User Focus
Mission Adoption • $400k-$1M • Mission Funding • Flight Readiness
Multi-Mission / Institutional • $1M - $5M • Multiple Mission • Becomes Heritage
• Missions ~receptive • Major bottleneck • Each mission different • Require extensive • Can try for AMMOS Tech “captain” of own ship validation for entry • Needed for approaching • Often afraid NASA won’t • Need to translate to missions like Data Science / ML “onboard” reqs • Try to navigate science-based • Fear becomes main R&TD process challenge Shared Repositories for past DS / Pot of Money to validate DS / ML Mission AO specifically requesting ML datasets & labels systems to mission ready status new DS / ML techniques 13
• JPL has this under control • R&TD system • Engineering Improvement • Data Science Working Group
14
2
Data Science Life Cycle Decision Autonomy Support
Archiving
Knowledge
Analysis
o Quick Analysis • Quick Calibration • Quick QC • Quick Analysis • Quick Knowledge o Annotate o Advise / React o Repeat
QC
Curation
Calibration
Operations
Transport
Acquisition
Formulation
Autonomy / Advisory Loop
Data Mining
Quick-Look products enable earlier reaction Advisory systems direct focus
3
What is Machine Learning? Algorithms that inductively self-assemble from examples. Expert Examples
Functional System
( e.g. Estimates 1-10 of attractiveness of picture )
• • • •
Training Data ( Encompasses domain of responsibility )
Interpretable as required Fixed code (V&V ready) Not ever-learning (mostly) EASILY UPDATED
Strength: Don’t specify rules Don’t explain “how” Machine Learning simplifies & systematizes the building and 4 updating of Autonomy / Advisory systems
Higher Level Questions / Actions “Where should I start looking?” “Show me more like this.” “What is likely to happen next?” “(Re)optimize my system.” ”How many kinds are there?” “Show me the most interesting first.” “What inputs are most informative?” “Show me new things.” 5
1. Science-driven use cases 2. Explanability 3. “Let Me Help” 6
Drs. Lukas Mandrake, Gary Doran, Brian Bue
Extremophiles
DHM field team Nuuk, Greenland
Raw Hologram (2D) Detections Contamination
• Digital Holographic Microscopes • Big data (4D, ~GB/s), rare findings • Motility ~ Life (composition agnostic!) • HELM ML system detects, tracks, and classifies in messy, raw 2D holograms
Global Surveys
7
Dr. David Thompson, Dr. Brian Bue, et al
CH4 detection in four corners
Enabled ground team to find underground pipe leaks
Airborne Imaging Spectrometer Multiple gas pipelines shut down / repaired Machine Learning “That Matters”
8
Dr. Lukas Mandrake Don’t Pre-Filter Data: No Data Quality Flags
instead
Order by Trust Learn from Experts
Current OCO-2 Quality Estimation Product / NASA Software of the Year runner-up
9
SMACC Endmembers Superpixel segmentation
Similar orbital tech used for • • • •
Minerology maps Identifying crop-types Recognizing diseased citrus Estimating hurricane damage
D. Thompson et al., 2012. TGARS.
10
• Borup Fiord sulfur springs • Biosignature analog site for Europa • Detect and track from orbit (EO-1) • •
Support vector machine (SVM) classifier 26 detections as of May 31, 2016 Courtesy: Steven E. Grasby
EO-1 image (Hyperion)
Sulfur detection (yellow) D. Gleeson et al., 2010. Remote Sensing of Environment. 11 L. Mandrake et al., 2012. ACM TIST.
12
Seedling Concepts • $30-$50k • Proof of Principle • Single Problem
Demo System • $150k-$300k • Extensive Validation • Single User Focus
Mission Adoption • $400k-$1M • Mission Funding • Flight Readiness
Multi-Mission / Institutional • $1M - $5M • Multiple Mission • Becomes Heritage
• Missions ~receptive • Major bottleneck • Each mission different • Require extensive • Can try for AMMOS Tech “captain” of own ship validation for entry • Needed for approaching • Often afraid NASA won’t • Need to translate to missions like Data Science / ML “onboard” reqs • Try to navigate science-based • Fear becomes main R&TD process challenge Shared Repositories for past DS / Pot of Money to validate DS / ML Mission AO specifically requesting ML datasets & labels systems to mission ready status new DS / ML techniques 13
• JPL has this under control • R&TD system • Engineering Improvement • Data Science Working Group
14
