|
|
 |
SENSITIVITY ANALYSIS OF ERS-1 SAR SIGNAL TO MULTISCALE-STRUCTURES OF
THE TROPICAL FOREST
BY MEANS OF THE WAVELET TRANSFORM
| Marc Simard, Gian Franco DeGrandi, Keith P.B.Thomson, Marc Leysen |
|
MTV-Space Applications Institute, Joint Research Center of CEE,Ispra,Italia.
Centre de Recherche en
G\'{e}omatique, Universit\'{e} Laval, Qu\'{e}bec,
Qu\'{e}bec, Canada.
simard@crg.ulaval.ca
|
|
Abstract
-
The work presented here is part of the TREES Central Africa Mosaic
project carried out in the Monitoring of the Tropical Vegetation Unit
of the EC DG JRC Space Applications Institute. TREES (Tropical Forest
Ecosystem Environment Monitoring by Satellites) in the current phase II
is an EC project funded by DGXI and coordinated by MTV SAI, whose main
goal is the set up of an operational tropical forest monitoring
system.
This paper will focus on the sensitivity of the \mbox{ERS-1} SAR
backscattering to structures in the tropical forest.
These structures are responsible for scene intrinsic texture in the image.
Since structures can exist at different scales, one needs a multiscale
approach in order to measure the resulting multiscale texture. The
analysis was done using a technique which decomposes an image into its
different scales in order to identify the scales at which observable
structures exist. In particular the technique is based on the wavelet
transform.
In order to identify observable structures in the primary tropical
forest at first speckle noise reduction is achieved using a technique
that preserves the original space resolution.
Once a structure has been detected and observed in ``low-noise''
SAR image, it is possible to verify their detectability on a standard
PRI format radar image. However the PRI images are heavily blurred by
the presence of noise. The multiplicative characteristic of the
speckle noise will modulate the image components at many scales. We
must therefore also consider speckle contribution when using the
wavelet decomposition for texture analysis. Such an analysis will allow
us to evaluate a detection threshold for structures and intrinsic
texture in the image.
Normalised scalograms or energy maps are built for different
scales from the wavelet
coefficients. It is seen that those maps bring complementary information
related to the spatial context, and can discriminate different classes.
It is also concluded that the multiscale texture are most readily
detectable at larger scales.
Keywords: SAR,texture,wavelet,speckle,tropical
Introduction
Since space-borne radar sensors provide an
all-time and all-weather surveying tool, they are ideal candidates for monitoring land cover when one
needs continental-scale and on-demand coverage. One such case is the
TREES Central Africa Mosaic Project (Malingreau, 1995), in the framework of which the research presented in this paper is in progress.
The mosaic is composed of over 450 ERS-1 images acquired over the entire bio-geographical domain of Central Africa. The images are 3-looks PRI format data with
a 12.5 meter
pixel size.
The mosaic geographic position straddles the equator and therefore the ecosystem is imaged both in the dry and the wet seasons at the same time.
Visual interpretation of the mosaic shows that a global classification is not
possible on an
intensity basis alone (Simard, 1996). Indeed, what we observe are in some cases
abrupt intensity changes
in similar cover types at the space-time boundary, namely at places where due to the satellite finite imaging time (two adjacent frames in longitude are acquired roughly within 20 days), the natural target has undergone radar cross-section
changes. We believe this
is most probably due to
meteorological factors such as rain and winds. Often, there is also a very
small difference of backscatter (lower than 1dB) for different targets,
such as primary and secondary forest.
Therefore, a contextual parameter has
to be defined.
We expect a measure of texture to bring new information.
But experience has shown that classical measures of texture such as
first order and 2-point second order statisitics, are useless for
classification purposes of ERS-1 SAR images of tropical forest.
From visual inspection
of the images, one can identify the large scale structures differences
between targets (see Fig.2). This hints to
multiscale texture analysis
of the SAR images.
This paper focusses on the use of the Wavelet Transform as
a contextual information extraction tool. The technique is very briefly introduced
in section 2. Then the transform is applied on the
SAR images and the results are discussed in section 3.
We then conclude in section 4.
THE WAVELET TRANSFORM
The Wavelet Transform is an extension of the classical Fourier transform where
a signal is decomposed onto an orthogonal basis. It is a mathematical
tool which allows for mapping of a signal onto a finite support basis.
The Wavelet Transform represents the original signal at different scales and
positions in space.
Because the basis has a finite support, contrarly to the Fourier transform,
one can obtain both frequency and
spatial information. The interested reader can refer to a tutorial paper by
Rioul and Vetterli (Rioul and Vetterli, 1991). The transorm allows for a multiresolution
analysis
of the signal and is a very interesting tool for multiscale analysis
of images.
The components of the signal at different scales and spatial locations
is computed from the
inner product of the signal I(x,y), and a wavelet $\psi(a,b)$ such that:
\begin{equation}
c_{a,b}=.
\label{eq:wavcoeff}
\end{equation}
where $a$ and $b$ are the scale and translation (position) parameters
respectively. $c_{a,b}$ is the wavelet coefficient and represent the
signal $I$ at a scale $a$ and position $b$.
The wavelets $\psi_(a,b)$ can be thought as a band-pass filter and can
be assimilated to a subband coding scheme. An efficient algorithm was
developed by Mallat (Mallat, 1989) for image analysis.
It consists of a series of high-pass and low-pass filtering steps
followed by subsampling. The algorithm results in the construction of
low-pass versions (lower resolution) of the original image and its components on the
wavelet basis. The wavelet algorithm also provides directionality
information, and in this way the image structures are represented according to
their orientation in the image plane.
APPLICATION OF WAVELETS TO THE SAR IMAGES OF THE TROPICAL TROPICAL FOREST
The Wavelet Transform is a relatively new tool for image analysis.
Few authors have reported about the Wavelet Transform of SAR images in the
international litterature (Du et al., 1993; Fau et al., 1994; Simard et al., 1997), eventhough it has been
extensively used for processing of other signal types and for
data compression.
The implication of SAR correlated multiplicative speckle noise has
been studied by Simard et al.(Simard et al., 1997), and it was concluded that
to avoid multiplicative
noise modulation for the analysis of texture, the
wavelet coefficients should be normalised. An example of the application
of this technique is illustrated next.
The original SAR image is shown on Fig.2. Two main classes are
of interest in this image because they cannot be distinguished by average
intensity alone: the degraded and primary tropical forest. The degraded
forest is composed of patches of forest, agricultural fields and small savannas.The primary tropical forest is a homogeneous region of dense vegetation.
The image was then
decomposed in the quadratic spline wavelet basis up to a scale of
200m with respect to the original 12.5m pixel size.
Scalograms or energy maps can be constructed for each scale. In order
to gather all available contextual information, we have added quadratically
the wavelet component of the three different orientations.
The resulting scalograms were then normalised by the low-pass version
containing average intensity information at resolutions lower than the
studied scale in order to avoid speckle modulation. Fig.3 shows
the
resulting energy maps which contain information due to intrinsic texture
of the scene. Contribution from noise is the same for all targets
independently from the average intensity. The algorithm is shown on Fig.1
The procedure was repeated for a ``low-noise'' and a ``noise-free'' image.
The latter was constructed from the quadratic average of 18 images of
a temporal series of the same scene,
without regards to average intensity changes. It gives only a way of evaluatingthe structures detectable by ERS-1 configuration and does not mean the same structures
can be found in a single PRI image. The ``low-noise'' image is constructed
from the quadratic average of 3 images, where the
average intensity was considered constant.
Such energy maps are presented
on Fig.3. Texture distinction between degraded (top left) and
primary tropical forest (center left) is very clear at a scale of 200m.
CONCLUSION
We conclude that texture exists at large scales in SAR images of the
tropical forest enventhough it is not measurable with classical techniques.
From a multiresolution analysis it is possible to
retrieve that contextual information which could then
be used for classification.
The wavelet decomposition provides a very efficient way of analysing
multiscale SAR image texture which is found at scales larger than a 100m.
From the energy maps of Fig.3, it is seen that the
structures are most
easily detectable for images with lowest noise, but a visual
inspection of Fig.3 also demonstrates the efficiency of
the algorithm
in extracting multiscale intrinsic texture information from
the standard ERS-1 SAR PRI images.
More work is in progress to characterise different targets (classes)
and understand the information extracted from the energy maps. That
new information will be at a later stage implemented in a
classification scheme.
Figure Captions
Fig.1: The algorithm for construction of energy maps. The first part
corresponds to Mallat's algorithm where the rows and columns are filtered
succesively with high-pass filter G, and low-pass H.
Each filtering is followed by subsampling of columns (Sc) and rows (Sr).
The detail images are quadratically added (+) and normalised ($\div$) by the
low-pass image to obtain the energy map (E) at a given scale.
Fig.2: (left) ERS-1 PRI image with a 12.5m pixel size and 25m resolution of
a tropical forest region located in Sassandra, Ivory coast. Noise is
reduced by quadratic
averaging of 3 (center) and 18 (right) PRI images of the scene.
Fig.3: Energy maps at a scale of 200m constructed from a raw PRI image
(left), from the quadratic
average of 3 (center) and
18 (right) PRI images respectively. It is seen that the
different targets such as the degraded forest (top left) and primary
tropical forest (center left of river) can be discriminated from their
texture content at
large scales.
REFERENCES
- J.-P. Malingreau and G. Duchossois,1995:
- The Trees/ ERS-1 SAR'94 Project,
Earth observation quarterly, ESA, 48, p. 48.
- M. Simard, F. DeGrandi, K.P.B. Thomson and G.B. B\'eni\'e,1996:
- Analyse multi-\'echelle de la texture de la mosaique ERS-1 de la for\^et
tropicale africaine, CDROM du $9^{e}$ congr\`{e}s de l'association qu\'{e}b\'{e}coise de t\'el\'ed\'{e}tection, Qu\'{e}bec.
- O. Rioul and M. Vetterli,1991:
- Wavelets and signal processing, IEEE SP magazine, 10, p. 14.
- S.G. Mallat, 1989:
- A theory for multi-resolution signal decomposition: The wavelet representation, IEEE Transactions on pattern and machine intelligence, 11, 7,
p. 674.
- R. Fau, G.B. B\'eni\'e, J.-M. Boucher and D.-C. He, 1994:
- Segmentation markovienne pyramidale d'imagesJournal canadien de teledetection, 20, 2,
p. 150.
- L.-J. Du, J.-S. Lee, K. Hoppel and S.A. Mango, 1993:
- Segmentation of SAR images using wavelet transform, International journal of imaging systems and technology, 4,
p. 319.
- M. Simard, F. DeGrandi, K.P.B. Thomson and G.B. B\'eni\'e, 1997:
- Analysis of Speckle Noise Contribution on Wavelet Decomposition of
SAR Images, Submitted IEEE Transactions on Geoscience and Remote Sensing
Keywords: ESA European
Space Agency - Agence spatiale europeenne,
observation de la terre, earth observation,
satellite remote sensing,
teledetection, geophysique, altimetrie, radar,
chimique atmospherique, geophysics, altimetry, radar,
atmospheric chemistry
|
EO Data Access
