319
RADIOMETRIC BLOCK-ADJUSTMENT OF SATELLITE IMAGES
REFERENCE3D® PRODUCTION LINE IMPROVEMENT
L. Falala 3, R. Gachet 3 , L. Cunin 3
a IGN, Institut Géographique National, 6 Avenue de l’Europe, 31520 Ramonville,
France - (laurent.falala, roland.gachet, laurent.cunin)@ign.fr
Commission, WG IV/3
KEY WORDS: SPOT, Radiometry, Mosaic, Block, Multitemporal, Orthoimage
ABSTRACT:
One layer of global geographic database Reference3D® is an image mosaic (5m resolution) made exclusively from panchromatic
SPOT5-HRS images. In order to improve Reference3D® production line, in particular SPOT5 HRS images mosaicking step, a new
strategy of radiometric block adjustment of satellite images was developed, implemented and tested at French mapping agency IGN.
By analogy with “geometric” block adjustment, we developed an iterative algorithm, adapted to Reference3D datasets (many
overlapping images covering a wide area), that calculates for each image a polynomial model to apply to its radiometry. These
polynomials are found by least-square resolution of a global linear system. Among many tests conducted to try different parameters
and validate the process, we present here results of a test case over Algeria (9 not too cloudy images taken in less than 3 months) and
another more difficult over Tasmania (11 very cloudy images taken in 4 years). In both cases, radiometric differences between
images were dramatically reduced. Reference3D contains also a DEM obtained by merging overlapping DEMs (calculated by
automatic image matching of SPOT5-HRS stereo-pairs). Radiometric block adjustment methodology can be easily adapted for
DEMs tilting, a preliminary step before merging. So, thanks to these new processes, previously manual tasks in reference3D®
production line are now mostly automatic.
1. CONTEXT
1.1 Reference3D® database
Reference3D® is a global geographic database produced by
Spot Image and French mapping agency IGN. It contains three
layers of information:
• DTED2 Digital Elevation Models (DEM with pixel
size at 1”),
• 5m panchromatic ortho-images,
• Quality and traceability metadata (including water
and cloud masks).
Reference3D® is made exclusively from SPOT5-HRS stereo
pairs with the following production line:
• Block adjustment of large sets of HRS stereo-pairs,
• DEM generation by automatic image matching of
stereo-pairs,
• Orthorectification of every HRS scene with
previously generated DEM,
• Image mosaicking and DEM merging,
• Quality metadata editing, tiling and packaging.
Annual production rate reaches 7 million km 2 in 2007 but
studies are conducted to improve productivity, especially to
automate previously manual tasks. For this purpose, a new
algorithm of automatic radiometric block-adjustment was
developed, implemented and tested on real Reference3D®
datasets. It performs automatic radiometric adjustment and
helps to build a cloud mask, which improve image mosaicking
and metadata editing steps. It can be used also to tilt DEM,
which may be necessary for DEM merging step.
1.2 Radiometric adjustment techniques
Radiometric block adjustment consists in modifying radiometry
of each image of a block in order to reduce radiometric
difference between images and create a more visually pleasing
and almost seamless mosaic. Various methodologies exist to
complete this task, each tailored to a particular context.
Radiometric adjustment, based on physical modelling of solar
illumination, has been developed for mosaics of cloud-free,
mono-date and multispectral aerial images (Martinoty 2005).
But Reference3D input images are multi-temporal and
distributed throughout the world, making any physical
modelling too complex if not impossible.
On the other hand, image harmonization and cloud detection
techniques exist for low-resolution satellite images with many
spectral bands, like NASA cloud masking algorithm on MODIS
images (Ackerman, 1998). But such methods can’t be used on
panchromatic SPOT5-HRS images containing only one spectral
band.
2. METHODOLOGY
2.1 Introduction
By analogy with “geometric” block adjustment, we developed a
radiometric adjustment methodology adapted to Reference3D
datasets made of many overlapping images covering a wide
area. This non-physical iterative algorithm calculates a
polynomial model for each image of the block, by least-square
resolution of a global linear system. At the end of this process,
radiometry of any pixel(colmun, row) of a given image I is
modified as follows (eq. 1):