DATA REDUCTION and ENHANCEMENT

1
DATA REDUCTION and ENHANCEMENT of GLOBAL
COMPOSITES of SPOT-VEGETATION (VGT)

• Herman Eerens, Else Swinnen, Yves Verheijen
• Vlaamse Instelling voor Technologisch Onderzoek
  (Vito - Belgium)
• Frank Canters
• Vrije Universiteit Brussel (VUB - Belgium)
• Acknowledgements
• Belgian Science Policy Office (Funding)
• JRC-SAI (Full year cycle of global VGT-S10)

2
MVC-Composites - still affected by clouds,
bidirectional effects, measurement errors - best
visible / removable in longitudinal analysis
(time series) - cleaning procedures MVC-month,
BISE, Verhoef,...
3
Extraction of Phenological Variables - Simple
Annual mean / min / max / amplitude of NDVI -
Complex start / end / length of green
season(s) - Often better inputs for
classification - Only feasible through
longitudinal analysis (time series)
4
Phenological Variables Examples from the Global
Land 1km AVHRR Data Set April 1992 - March 1993,
Interrupted Goode Homolosine
5
Phenological Variables Examples from the Global
Land 1km AVHRR Data Set April 1992 - March 1993,
Interrupted Goode Homolosine
6
Phenological Variables Examples from the Global
Land 1km AVHRR Data Set April 1992 - March 1993,
Interrupted Goode Homolosine
Simple Biome Classification Bivariate
Level Slice - Annual Mean NDVI -
Annual NDVI Range
7

• LONGITUDINAL TIME SERIES ANALYSIS REQUIRED
• It adds a new dimension to the results of the
  transversal analysis (per decade / day)
• At a given moment, áll information of a full
  year (36 x S10, 360 x S1) must be available
  simultaneously

Band BPP CONTENTS VGT-SYNTHESIS ------- ------
----------------- 1 2 BLUE Reflectance 2
2 RED Reflectance 3 2 NIR Reflectance 4
2 SWIR Reflectance 5 1 NDVI 6 1 Zenith
angle of sun 7 1 Zenith angle of
sensor 8 1 Azimuth angle of sun 9
1 Azimuth angle of sensor 10 1 Status map
errors in 4 bands, land, cloud, snow/ice 11
2 Time grid minutes between pixel registration
and start of synthesis (LOG-file) Total 16 GLOB
AL VGT-SYNTHESIS ? 600 Mb pixels x 16
byte/pix ? 10 000 Mb ? 10 Gb FULL YEAR
CYCLE S10 360 Gb S1 3.6 Tb

• TOO MUCH DATA ? CLASSICAL SOLUTIONS
• Temporal selection select limited number of
  composites
• Spectral selection only NDVI,
• Spatial selection - Extract/analyse specific
  study areas
• - Work on degraded images

8
DEDICATED SCHEME FOR DATA REDUCTION
1. Radiometric Compression 16 ? 6 bytes
(37.5) - Eliminate BLUE and NIR - Rescale
reflectances from 16 to 8 bit (0-250) RED/SWIR
R 062.5 in steps of 0.25 NIR R 083.3
in steps of 0.33 - Values 251-255 special
flags (saturation, error, ...) - Status_out
Cloud Snow/ice Day_in_decade - Combine 2
zenith angles in 1 byte (steps of 5) - Combine
2 azimuths in 1 byte (relative azimuth0-180)
? Output 6 Byte-layers 2. Eliminate all
the Water Pixels (25) - 134,134,736 land pixels
left ? Spatial context lost! - Results stored
in Pseudo-Images (PI) ? IDL/ENVI-images
(Ncol5000, Nrec26827) - All spectral
(per-pixel) operations still feasible (via
IDL/ENVI) time series analysis,
classification (!), ... 3. Improved
Land/Sea-Mask - VGT-mask 5-10 sea pixels along
coast (too much) - Boreal regions in winter
confused with sea (status map) 4. Conversion to
Equal-Area Projection (IGH) - In Plane
carré of VGT (worst projection) at (1/112)² -
Out Interrupted Goode Homolosine (IGH) at
1x1km² ? REDUCTION 10 without LOSS of DATA
! ? FULL YEAR CYCLE S10 36 Gb S1 360Gb
9

10
Example of IDL/ENVI-FORMATTED PSEUDO-IMAGES
NORMAL IMAGES A Land-See Mask (Inter. Goode
Homolosine, 1x1km²) B GTOPO30-DEM (Geographic
Lon/Lat) C VGT-S10, Dec.3 of May 1998, NIR
(Geogr. Lon/Lat) PSEUDO-IMAGES (only 134 135 000
land pixels) 1 Longitude of pixel centre
(float) 2 Latitude of pixel centre
(float) 3 Altitude (from B - Short Integer) 4 NIR
of Dec.3 of May 1998 (from C - rescaled to Byte)
11
Example of IDL/ENVI-FORMATTED PSEUDO-IMAGES
1. PI_XY IN Land/Sea mask in Master System
(IGH) OUT 2 master-PIs with IGH-X/Y of pixel
centres 2. PI_IGH IN 2 master-PIs with
IGH-X/Y of pixel centres OUT 2 master-PIs with
Lon/Lat of pixel centres 3. PI_EXTR IN Any
image in IGH or Lon/Lat ( master-PIs) OUT PI-v
ersion of that image (Byte / Short Int /
Float) 4. PI_VGT IN VGT-S10/S1
master-PIs OUT 6 Byte PI-images 5.
PI_BACK IN Any previously created PI (
master-PIs) OUT Corresponding normal image in
IGH Option selection of output window 6.
PI_REDU IN Set of all VGT-PIs (
master-PIs) OUT Corresponding set of normal
images, IGH, degraded resolution (33x33km²),
systematic selection 7. CLEAN IN Set of 36
VGT-S10 images (normal or PI) OUT Cleaned
NDVI-profiles 8. PHENO IN Set of 36 VGT-S10
images (normal or PI) OUT Cleaned
NDVI-profiles
12
SET of DEGRADED IMAGES
RED NIR SWIR
End of June 1988
NORMAL IMAGES - 36 decades x 6 216 images -
Global but degraded (33km x 33km) - Npix
1213x423 513 099 ? Total 216 x 0.5 110 Mb -
Systematic pixel selection ? original
signatures - Excellent data set to test
performance of new procedures on global scale
13
CONCLUSIONS
1. One Possible Pathway for Global
Classification - Transverse reduction of all
VGT-images ? PIs - Also extract external
information DEM, ? additional
classification variable Regions ? for
post-processing (LC-statistics) Existing
classifications ? Training / Validation -
Longitudinal analysis on PI-images Cleaning,
elimination of bidirectional effects, addition
of phenological variables, improved VIs, -
Classification in PI-form - Reconversion to
normal image 2. Preliminary data enhancement via
data reduction seems indispensible 3. To be
integrated in CTIV (?) ? Optional delivery of
data in PI-form (better than ZIP) ? More users
get access to global data 4. Lots of improvements
possible other geo-systems, other output
formats (now only ENVI IDRISI), streamlining
of software, 5. High radiometric resolution
redundant