Virtual Sensors

1
Virtual Sensors

• Nikunj C. Oza, Ph.D. oza_at_email.arc.nasa.gov
• Member, Discovery, Prognostics, and Adaptive
  Systems Group
• Discovery and Systems Health
• Joint work with Ashok Srivastava, Ph.D.
• Julienne Stroeve, Ph.D. (NSIDC, University of
  Colorado)

2
Outline of Talk

• Motivating problem---detecting clouds over snow
  and ice.
• General Virtual Sensors problem.
• Results.
• Related accomplishments.
• Summary.

3
Cloud Detection back in Time

• Needed to retrieve surface geophysical parameters
  such as albedo and temperature.
• Difficult over snow, ice-covered surfaces---low
  contrast in visible and thermal infrared.
• Terra and Aquas MODIS 1.6mm has enough contrast
  for this task.
• Problem Only available since 1999.
• AVHRR/2 in operation since 1981.
• However, 1.6mm channel not available in AVHRR/2.

MODIS Spectral Measurements
AVHRR/2 Spectral Measurements
No Black Signal for AVHRR/2. Need to emulate
AVHRR/2 Cloud Signal
4
Cloud Detection back in Time

• Solution Predict 1.6mm channel using a Virtual
  Sensor

MODIS Spectral Measurements
AVHRR/2 Spectral Measurements
Black Signal for AVHRR/2 estimated by a Virtual
Sensor.
5
Virtual Sensors Approach

• Given
• MODIS channels 1, 2, 20, 31, 32 correspond to
  five AVHRR/2 channels
• Develop
• Model MODIS channel 6 (1.6mm) as a function of
  five MODIS channels
• Apply
• Use function to construct estimate of 1.6mm
  channel for AVHRR/2

model
MODIS 1,2,20,31,32
MODIS 6
Model Construction
model
AVHRR 1,2,3,4,5
AVHRR 6
Model Application
6
Virtual Sensors

• Virtual Sensors predict the historical record for
  spectral measurements using relationships found
  from existing sensors and inputs from historical
  record.
• Useful for simulating sensors back in time

Accuracy of learned models for MODIS data
70-90 (over 2 weeks)
Channel 6
MLP prediction
SVM models have found different melt regions
7
Accuracy Results for Three Models
True Positive
True Negative

• True Positive number of times channel 6
  indicated a cloud and the model predicted cloud
• True Negative number of times channel 6
  indicate no cloud and the model predicted no cloud

8
Utilize Kernel Methods to improve performance

• We build the Virtual Sensor using new
  mathematical techniques known as kernel methods.
• Map the data using a function F(x) into a high
  (or infinite) dimensional feature space.
• Procedure
• Map the data into a high (or infinite)
  dimensional feature space
• Perform linear operations in feature space (e.g.,
  discriminant analysis, regression)
• Map result back to the original space
• Caveat All mathematics must use inner-products

9
The value of kernels
Original Data without Kernel
Mapped Data using Kernel
F(x) Kernel Map
Data in original space highly complex decision
boundaries.
Data in high dimensional feature space after
mapping through F(x) can yield simple decision
boundaries.
10
Kernel Method Example Support Vector Machines

• Solve convex optimization problem gt one globally
  optimal solution.
• Mitigate overfitting
• e-insensitive loss
• Allows errors but bounds them

11
Verification of Models on MODIS Data
12
Application of Models to AVHRR Data
13
Significant Improvements in Processing Times

• At the National Snow and Ice Data Center
• Built and deployed emulated physics models to
  calculate corrections for surface albedo
  measurements
• Deployed data mining algorithms that increase
  processing speed by factor of 27 compared to
  existing DAAC methods
• Potential to deploy Virtual Sensors for
  generation of a historical cloud mask record

14
Summary

• Virtual sensors enable retroactive long term
  climate models that would otherwise be impossible
  to obtain.
• Benefits
• Increases return on investment in new
  instruments measurement capability by extending
  it back in time.
• Facilitates long-term analysis, climate modeling.
• Can guide use of expensive, high-resolution
  instruments.
• Model Predict high-resolution measurements and
  confidence given low-resolution measurements.
• Use high-resolution instrument only when
  confidence is low.