Machine Learning and Instrument Autonomy Group

Mission

To research, develop, and infuse rigorous solutions built with applied Machine Learning and related technologies to support the exploration of Earth & Space and the advancement of science.

Infusion Targets

▶ Missions
      ▸ Earth: GRACE-FO, OCO-2/3, EO-1, ASO, AVIRIS-NG, VCAM, MISR, NISAR, SWOT, SMAP, MODIS, GOSAT, AIRS
      ▸ Cubesats: IPEX, NEAScout
      ▸ Rovers: MER, MSL, M2020
      ▸ Deep Space: Europa Clipper, Europa Lander, MRO, Cassini
      ▸ Ground: V-FASTR, PTF, ZTF, DSN, VLBA, VAST, CAST, TCCON, DoD

OCO-2 Spacecraft ▶ Archival
      ▸ Focus of Attention / Content-based Search
      ▸ Validation, Uncertainty/Quality Estimation
      ▸ Data Exploration & Mining

▶ Real Time / Operational Pipelines
      ▸ Focus of Attention / Rapid Response
      ▸ Triage / Prioritization
      ▸ Quality Estimation

▶ Anomaly Detection / Root Cause Analysis
      ▸ Onboard / Autonomy
      ▸ Triage / Prioritization
      ▸ Knowledge Compression / Summarization
      ▸ Auto-calibration
      ▸ Change Detection
      ▸ Reactive Systems

Hyperspectral Data Cube

Areas of Expertise / Breadth

▶ Diverse data domains
      ▸ Atmospheric Composition (GCMS)
      ▸ Exoplanet, Atmospheric, Planetary Surface, Near & Deep IR Spectroscopy
      ▸ Orbital (Nadir/Limb) Imagery Earth/Planetary
      ▸ Terrestrial scatterometry / altimetry
      ▸ Orbital/Airborne Synthetic Aperture Radar (SAR)
      ▸ Astrometric, Light Curves
      ▸ Atmospheric/Ionospheric Soundings
      ▸ Hydration via Neutron & Electrochemical Impedance Spectroscopy
      ▸ GPS
      ▸ Langmuir & Plasma Spectrometry
      ▸ Mission telemetry

▶ Data formats
      ▸ Images, Hyperspectral Images, Image Time Series
      ▸ Spectral & Spectrographic
      ▸ High-dimension multivariate time series
      ▸ Geospatial & Spatiotemporal

▶ Technologies & Models
      ▸ Unsupervised (Clustering, Segmentation, Autoencoding, GAN)
      ▸ Supervised (Logistic, Decision Tree, SVM, Random Forest, LSTM, CNN)
      ▸ Classification & Regression
      ▸ Dimensionality Reduction / Visualization (PCA, ICA, MNF, TSNE, SVD)
      ▸ Ensemble (Voting, Boosting)

GEOS-5 Model Assimilating OCO-2 Data

Specialization / Depth

▶ Explainability / Hypothesis Generation
      ▸ Most models built for understanding, not automation
      ▸ Feature selection / alternative feature discovery
      ▸ Provide justification for classification / selection
      ▸ Provide user guidance to help build data intuition
      ▸ Feature engineering to encapsulate learned data structure

▶ Model Uncertainty Quantification
      ▸ Empirical Probability of Correct Classification (not just confidence)
      ▸ Analysis of regions / subsets with high / low performance

▶ Rigorous Verification & Validation
      ▸ Cross-validation of all results, including deep learning
      ▸ Broad sweep hyperparameter optimization
      ▸ Stability analysis

HiRISE images of Seasonal Flows on Warm Martian Slopes ▶ Domain-relevant Impact Metrics
      ▸ Use-case requirement driven
      ▸ Speed of data exploration (hours)
      ▸ Speed of anomaly resolution (hours)
      ▸ New mission capabilities / concepts enabled
      ▸ Reduced data downlink requirements (%)
      ▸ Lowered operational costs ($)
      ▸ Reduced operator skill required

▶ Physical Science Support
      ▸ Publication-supporting product quality
      ▸ Explainable, justifiable model behavior
      ▸ Rigorous performance statistics
      ▸ Intuition and interpretation facilitation
      ▸ Focus-of-attention to permit manual inspection and verification

▶ Model Compression
      ▸ Limited compute environments (onboard autonomy)
      ▸ Low latency / real-time requirements (Ops / autonomy)
      ▸ Multi-model performance comparison / Data insight generation
      ▸ Quick-look capabilities before large analysis computation investment
      ▸ Enhanced explainability / Unnecessary parameter reduction

For more information about working with our group, please contact Lukas Mandrake.

Group Photo