MODIS and VIIRS Geolocation Error and Long-term Trend Analysis with Automated Correction Techniques using Kalman Filtering Masahiro Nishihama Sigma Space Corporation Lanham, Maryland, USA (20706) [email protected]Robert E. Wolfe NASA Goddard Space Flight Center, Code 614.5 Greenbelt, Maryland, USA (20771) [email protected]IGARSS 2011 July 24-29, Vancouver Canada
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
MODIS and VIIRS Geolocation Error andLong-term Trend Analysis with Automated
INTRODUCTION - Accurate geolocation of remote sensing data is needed for Earth science
research and applications
- Initial on-orbit bias removal is performed immediately after launch
- Analysis and removal of long-term geolocation bias is needed to maintain accurate geolocation throughout multi-year missions
- Analysis and removal of the within-orbit trend is also needed because of thermal effects on instrument pointing
- Current approach is a manual least-squares analysis that uses linear and sinusoidal curves to remove both the long-term and within-orbit trends
- This presentation discusses two Kalman filtering approaches that are being considered to replace the current manual least-squares approach
- These new techniques can be applied to the Moderate-Resolution Imaging Spectroradiometer (MODIS) instruments and the future Visible Infrared Imaging Radiometer Suite (VIIRS) instrument.
IGARSS’11-Nishihama et al. 2
Background• MODIS - Launched on NASA’s Earth Observing System (EOS) Terra spacecraft in December 1999, and started collecting data in February 2000 - MODIS on Aqua spacecraft in May 2002, began collecting data in July 2002 - whisk-broom sensor with 36 spectral bands; 2 at 250 m, 5 at 500 m and 29 at 1 km nadir spatial resolution - “Ideal” band is located at the center of focal planes (1km resolution)
• Simulated MODIS-like Image with TM scene and MODIS Geolocation file - Use triangular weights and assemble TM samples to build a MODIS like pixel with 250m resolution
MOD02QKM.A2001199.0840.004.2003121094844.hdf 2001 Day 199 -- July 18 From TM: L5176039_03919840920_B10.TIF
L1B: 8160 lines x 5416 samples
IGARSS’11-Nishihama et al. 6
• Geolocation Residuals to Statistical Analysis - Convert along-scan/track shifts to Nadir view angle offsets- Convert them to ground distance and display by Scan Angle.
IGARSS’11-Nishihama et al. 7
• Long Term Geolocation Error Trend and Analysis - One year or more of accumulated residual records needed - Trend (linear and annual cyclic) with a combination of linear and sinusoidal curve to fit the scan and track directions - Remove additional smaller within-orbit variations caused by temperature variation - Reduced geolocation error to less than 50 m Root Mean Square Error (RMSE) for MODIS/Terra and 60 m RMSE for MODIS/Aqua.
Note: Difference in Geolocation Accuracy between Terra and Aqua
- Terra orbit is identical to Landsat orbit,
- Terra observes a similar Earth shadow as TM does (10:30 AM) - Motion of AMSR-E Parabolic Reflector 1.6 meters in diameter on top of Aqua
IGARSS’11-Nishihama et al. 8
• Terra/MODIS Adjusted Residuals - along Track and Scan
-75
-50
-25
0
25
50
75
0 1 2 3 4 5 6 7 8 9
Tra
ck (a
dj.)
re
s. (
m)
.
Years since Jan. 1, 2000
Daily 16-day Global 16-day Southern Hemisphere 16-day Northern Hemisphere
-75
-50
-25
0
25
50
75
0 1 2 3 4 5 6 7 8 9
Sca
n (
ad
j.) r
es.
(m
) .
Years since Jan. 1, 2000
IGARSS’11-Nishihama et al. 9
• Two Kalman Filtering Approaches for Long-term Trend Analysis 1. Linear and Sinusoidal Filter - Replace the manual approach with the automated approach by a combination of linear and sinusoidal curves with a period of one year, day-by-day, - Estimate roll/pitch errors to the instrument alignment matrix to new geolocation ,
2. Satellite Attitude Error Estimation Filter: - Express the attitude variation(error) as differential equations of Euler angles and use residual errors as part of observations
- Estimate residual Roll/pitch/yaw error to alignment matrix correction to new geolocation .
IGARSS’11-Nishihama et al. 10
• Linear Kalman Filter Key Parameters and Equations
- Definition of Time steps t0, t1, t2……. (days)
- Non-Linear Differential Equations to Linear Equations
- Observations or Meaurements z …… (along-track/scan residuals) - Initial State Vector X0
- System Noise (Q) and Measurement Noise (R) Covariance matrix - Initial State Error Covariance Matrix P0 ,
- Fundamental Matrix (Φ), Derived at tn+1 from tn
- A priori State Error Covariance Matrix (P-) and a priori Vector X-
- Kalman Gain Matrix (K) - Posteriori Error Covariance Matrix (P+) and best estimate X+
HXz
wAXX
IGARSS’11-Nishihama et al. 11
• Method 1. Linear and Sinusoidal Approach with
txtxtxxtx dcba cossin)(
dd
cc
bb
aa
dcb
x
x
x
x
txtxxx
0
0
0
0
sincos 0
)()( twXtAX
00001H
dcba xandxxxx ,,,,where State vector X:
where x(t) is either along-scan or along-track error with observation equation
• Preliminary Results- From track/scan residuals over 10.4 years
IGARSS’11-Nishihama et al. 20
DISCUSSION
1. Analysis of both Kalman filtering methods are preliminary, but results of the first method is very promising.
2. The second method needs more tests and examination. Appearance of yaw in the second method is interesting and needs investigation.
3. Further tests will be needed for Aqua data using both methods.
4. Determination of the initial covariance matrices and other parameters is critical, and requires careful attention.
5. The Kalman filters methods are computationally efficient, it took very little time to processing 3800 sample points (~250 ground points are acquired each day).
6. Once fully tested, either of the methods could be implemented, enabling automated daily geolocation parameter updates.