Definition and Validation of Scientific Algorithms for the SEOSAT/Ingenio GPP

Definition and validation of scientific algorithms for the
SEOSAT/Ingenio GPP
Eduardo de Miguel1, Raúl Valenzuela2, Teodoro Bernardino2, Verena
Rodríguez2, Alberto Pizarro3, Diana de Miguel3 , Severino Fernández1
Definition and validation of scientific
algorithms for SEOSAT/Ingenio GPP
ESRI users conference, Madrid, Sept. 2011
1 Área de Teledetección, INTA. Cta. de Ajalvir s/n. 28850 Torrejón de Ardoz, Madrid (España). Correo
electrónico: demiguel@inta.es
2 GMV Aerospace and Defence, S.A.U. C/ Isaac Newton, 11. P.T.M. Tres Cantos, 28760 Madrid
(España).
3 EADS-CASA Espacio. Avenida de Aragón 404, 28022, Madrid (España).

In the framework of the Spanish Earth Observation mission Ingenio, a
Ground Processor Prototype is being developed by GMV under contract
with EADS-CASA Espacio.
This GPP is part of an end-to-end mission simulator, aimed to:
1) be used for mission performance analysis during development and
verification phases,
2) be the basis for the future operational L1 ground processor, to be
Background (I)
2) be the basis for the future operational L1 ground processor, to be
implemented within the Payload Data Ground Segment.
Área de Teledetección - INTA supports GMV in the definition of the GPP
scientific algorithms and is responsible for their validation.

Sun-synchronous orbit, 678 km, repetition cycle 49 days.
Twin cameras, FOV ≈2.5º each; total swath 55 km.
4 bands VNIR 10 m GSD; 1 pan band from 0.6 to 0.8 µm, 2.5 m GSD.
Panchromatic band uses a 7 stages TDI.
minimum MTF@Nyquist = 0.11 (PAN).
Spectral separation with filters.
Background (II): Main characteristics of Ingenio
Spectral separation with filters.
Pointing accuracy:
Along track 500 m
Across track 100 m
Onboard pre-processing: equalization and dark signal removal.
Lossy compression applied to scientific TM.

The GPP includes the following tasks within the SEOSat/Ingenio L1 processing
chain:
•identification of saturated and bad pixels,
•radiometric correction and calibration,
•image restoration (denoising and deconvolution), initially only for Pan band
(TBC),
Background (III)
•geolocation and estimation of Rational Polynomial Coefficients (RPCs),
•basic image classification (land / sea / bright),
•quicklooks generation.

The algorithm selected is based on wavelet transform thresholding.
Why wavelets?
1) Noise is white (= equal power for all spatial frequencies),
2) Noise level is known a priori:
-instrumental noise from system characterization and in orbit test
-photon noise from the signal level at each point
L1b processing: denoising
3) Limitations in other noise-removal methods (median filter, Fourier
Transform...)

Algorithm configuration.
•nominal wavelet basis: symlet order 4
•up to 5 decomposition levels
•noise estimation at each level from the expected instrumental and
photonic noise (3sigma):
Noise (i,j) = 3 * sqrt(DarkN*DarkN + PhotoGain*DC(i,j))
DarkN: the instrumental noise in digital counts
L1b processing: denoising
PhotoG: the number of photons mapped into a digital count
• 75% shrinkage of those coeff<noise
• proportional reduction of the other coefficients (soft thresholding)
Important: noise estimation based on theoretical values for DarkN y PhotoG.

The first step has been building a rigorous model of system
signal and noise for each spectral band.
This model has been applied to a synthetic scene and to a
airborne image acquired and processed by INTA.
Denoising validation
Lref
Lmin
Lmax

Denoising results
Nominal case

Denoising results
PhotoGain effect
Results on natural scene

L1b processing: deconvolution
Inverse filtering using a specific kernel.
The kernel is derived by inverting the expected instrument PSF.
This procedure is a Wiener
filtering if noise is negligible.
The expected PSF is an input => must be evaluated continuouisly during the
mission

Deconvolution validation
The validation is based on analysis on the results over a synthetic
scene and a natural scene.
The synthetic scene is a bar pattern with spatial frequency 0.5
cycles/pixel. The scene is convolved with the expected Ingenio PSF
and noise is added. Next, denoising and deconvolution are applied and
the final contrast is tested:
SW/SSD <1.1 after denoising and deconvolution => MTF@Nyquist >
0.35.
On the natural scene the procedure is similar, but we look for artifacts
and not for a quantitative contrast verification.

Deconvolution results

Deconvolution results
Imagen AHS original
Imagen degradada
Wiener
Diff rest - original
Min: -79
Max: 82
Media: 0.117
Desvstd: 1.793

The geolocation procedure is challenging due to:
-the accuracy goal requires the use of GCPs,
-focal plane with different detectors, each one with a different line of sight
along-track.
Geolocation (I)

-Filtering of raw AOCS data to reduce noise.
-Initial geolocation using a rigorous instrument model and the filtered AOCS info.
- Automatic selection of GCPs on a reference image (GMV tool) for:
-estimation of thermoelastic bias
-estimation of roll and pitch AOCS error per image line by fitting the errors
per GCP to a continuous curve.
Overview of the geolocation procedure
Geolocation (IIa)
per GCP to a continuous curve.
-Fine-tunned geolocation using the bias and roll-pitch corrections.

Resample of the geolocated pixels on a UTM grid:
-sampling interval 2.5 m / 10 m
-inverse geolocation using High-order B-spline resampling.
Simultaneously, a "perfect sensor" image (Pleiades way) will be produced.
RPCs will be produced for this product.
Geolocation (IIb)
Overview of the geolocation procedure (cont.)
RPCs will be produced for this product.

The algorithms in the GPP shall allow the geo-location accuracy of L1b data
wrt the reference ellipsoid to be better than 50 m at 2σ without GCPs.
The GPP shall allow the geo-location accuracy of L1c data wrt a reference
map to be ≤ 2.5 m at 1σ with the use of < 20 GCPs
Geolocation validation: which is the requirement?
Geolocation (III)
The sum of the errors due to implementation of the GPP algorithms for the
geolocation at level 1C with GCP should be less than 0.1 pixels RMS (0.3
pixels, max. value).

The geolocation algorithm is validated using an independent geolocation
model developed at INTA. This is based in the ESA Earth Explorer C
libraries (http://earth.esa.int/cfi/)
The geolocation adjustment with the use of GCPs has been validated in
two steps:
Geolocation Validation
Geolocation (IV)
-capability to detect enough valid GCPs.
=> Test passed for nominal situations (the deformation between the raw
and reference images is limited, no clouds nor water bodies...).
-capability to adjust roll and pitch errors from GCPs offsets
=> Test passed for nominal scenes, and excluding "extrapolation"
segments at the scene borders.

EEtool vs GPP (initial result)
Geolocation (Va)

RMS error for a complete Ingenio scene vs number of GCPs
Geolocation (Vb)

Conclusions
The validation of image restorarion algorithms confirms they are suitable
for the mission. Moreover, they are robust against uncertainties in the
parameters involved in the image generation (PSF, DarkNoise,
PhotoGain...)
The algorithms selected for image geolocation achieve the required
accuracy in nominal conditions.
The validation of the scientific algorithms for different scenarios and
The validation of the scientific algorithms for different scenarios and
including non-nominal conditions is a challenging task.
The definition, implementation and validation of the GPP algorithms is a
joint work from by many engineers at GMV, EADS CASA Espacio,
Astrium SAS and INTA, with the continuous support of ESA Ingenio
project office and CDTI and IMAG (Ingenio Mission Advisory Group).

Definition and Validation of Scientific Algorithms for the SEOSAT/Ingenio GPP

More Related Content

What's hot (6)

Viewers also liked (8)

Similar to Definition and Validation of Scientific Algorithms for the SEOSAT/Ingenio GPP (15)

More from Esri (20)

Recently uploaded (20)

Definition and Validation of Scientific Algorithms for the SEOSAT/Ingenio GPP