You are on page 1of 3

I"I.

kNCE

OF SPECKLE FILTERING ON THE ASSESSMENT

OF AGRICULTURAL SURFACE PARAMETERS

S,FOUCHER "), F. ZAGOl,SKI'2', C . GAILLARD"),


Y. VOIRIN"' and M.N. NGUYEN'"
(1) Centre dApplications et de Recherches en Ttltdttection, (CARTEL - Universitt de Sherbrooke),

Phone: (+1) 819-821-8000 Ext. 2975


(+1) 819-821-7944

2500 Boulevard de l'Universitc-5, Sherbrooke (Qc) - CANADA J1K-2R1.

(2) PRIVATEERS N.V., Private Experts in Remote Sensing, De Weaver Drive 42, Philipsburg, St Maarten - Netherlands Antilles. c/o 228 Gillespie street, Sherbrooke (Qc) - CANADA J1H-4x3.

w:sfoucher@callisto.si.usherb.ca
ABSTRACT:Retrieval information about soil moisture and roughness from remotely sensed data is of great interest for agricultura4 soil management. This is particularly true for the soil c o m a t i o n programs. In fact, these two physical parameters are crucial for predicting erosion rates within agricultural watersheds during the rainfall season. Xn this context, there exists a strong interest to assess the potential of S A R (Synthetic Aperture. Radar) images to provide information regarding to these two physical surface parameters with periodic acquisitions. Nevertheless, one major source of error in the quantitative estimate of these parameters is the presence ctf the speckle within the scene registered by SAR instruments. To overcome this difficulty many filtering techniques n developed for reducing this multiplicative noise. as to our knowledge, no study has analysed the sensitivity of these speckle filterering
influence of seve assessment of these ering methods on the 'Filtering methods', allow to attenuate speckle fluctuations thanks to spatial averaging. Both methods enhance radiometric resolution at the expense of spatial resolution. Speckle filtering of SAR images is a primordial step in extracting the useful signal (Le., the underlying scene radar reflectivity or backscattering coefficient, 0") to be inverted for retrieving information about physical characteristics of the target. Although these filtering approaches have been tested with image processing criteria, such as the preservation of edge gradient value and the smoothing degree of homogeneous areas, a main issue is to their understand impact on the assessment of physical surface parameters. In this paper, we attempted to quantify the SAR filtering impact on the estimate of soil moisture and roughness on an experimental agricultural site in Normandy (France). This study was conducted using RADARSAT-SGC time-series for which a set of filtering techniques (i.e., 'box', 'median' filters, and more sophisticated filters using wavelet representation and simulated annealing technique) was applied. Results stressed the significant impact of S A R filtering technique in the assessement of soil paremeters, especially in the case of the soil roughness.

INTRODUCTION
One key issue in SAR (Synthetic Aperture Radar) remote sensing data exploitation is the strong radiometric variability of extended surface targets due to the presence of a signaldependant multiplicative noise, so-called 'speckle'. This noise directly results from the coherence of microwave radiations which induce unpredictable interference phenomena within SAR cell resolution. This makes the visual interpretation of SAR images and the performance of all the statistic based processing techniques such as image classification and segmentation very difficult. Many techniques have been developed to attempt to reduce speckle within SAR images. A commonly used approach, namely 'Multilooking technique', consists in applying an average of several single lock images. Other techniques, the

SPECKLE REDUCTION TECHNIQUES


Specke filtering methods can be separated into two categories: the non-adaptive approaches (box filter, median,. ..) and adaptive techniques. Speckle reduction can be achieved using a simple box filter which enhances radiometric resolution without any consideration on the target nature. Consequently, edges and other significant details are strongly degraded. The median filter allows to better preserve edge properties but with a possible bias in the radar reflectivity estimation. By using adaptive filters a compromise exists between the radiometric enhancement in the homogeneous areas and the preservation of the spatial resolution within the heterogeneous areas (i.e., the textural

0-7803-5207-6/99/$10.00 0 1 9 S IEEE.

2137

area or edges). Based on local statistic estimations, these filters compute the local normalized standard deviation value of the observed intensity which can be related to known speckle statistics. Depending on the local heterogeneity degree of the target, pixels are weighted with a value ranging from the local mean to the raw intensity values. Filters differ from the local weight determination depending on the APriori hypothesis applied on the probability density function (pdf) of both speckle and radar reflectivity. Thus, filters can be distinguished by the estimation strategy used such as the Minimum Mean Square Error (MMSE) or the Maximum A Posteriori (MAP). As an example, whereas Lee's filter [ l ] assumes gaussian pdf and apply a MMSE criterion, GammaMAP filter [2] assumes gamma pdf and uses a local MAP estimation. Oliver proposes a global MAP estimation based on the Metropolis algorithm [ 11. Integration of both optimal target and edge detection with an adaptive size window strategy can improve filtering results [2]. Wavelet filters are based on a multiscale representation of the image where high frequencies (Wavelet coefficients) are denoised using a MAP criterion and gamma pdf assumptions [3].

30" and 65". For each of the selected samples a least-square fit to (1) was applied on the 4 corresponding angular SAR backscattering coefficients for assessing E and (k.h). This was performed for each set of angular SAR data (i.e., 0" intensity resulting from parcel averages and 0"estimated by each of 5 filtering). Samples with a too small pixel number (~121) and for which estimates were unrealistic ( K O and/or k.hcO) were discarded from our analysis. A degree of heterogeneity level for each parcel (sample) was evaluated using the normalized standard deviation coefficient of the underlying radar reflectivity (Cd. This allowed one to define two groups of 59 samples: (a) homogeneous and slightly heterogeneous parcels (OICi<O. 12); (b) heterogeneous parcels (ci20.12).

RESULTS
Figure 1 displays, for each of two sample groups (i.e., homogeneous and heterogeneous parcels), the mean value of correlation coefficients <r> between 0 " measured with ' values each beam mode and corresponding fitted 0 derived from (1) using each of 5 filtered SAR images and the parcel average. The high values of <r> (- 0.7) confirms that filtered images with all speckle reduction methods are physically meaningful. As an example Figure 2 illustrates a case of correlation for the Gamma-MAP filter.
1.0

DATA SET
RADARSAT-SGC time-series (4 dates) were acquired over an experimental agricultural site in Normvdy (Blosseville, France) under several incidence angles (standard S4, S5, S6 and S7 beam) during March 1998. These SAR data were over-sampled, reducing the 25 m nominal resolution to 12.5 m, with a number of looks around 4. Moreover, a SPOT-XS image from 1997 completed our remote sensing database.

........................................................
El Homogeneous parcels
Heterogeneous parcels .......................................................

g .E

METHODOLOGY
Five speckle reduction methods were applied on all the S A R images using the following calibrated specifications: (1) box filter (with a window size of 11x 11 pixels); (2) median filter (5x5); (3) Gamma-MAP filter (11x11); (4) simulated annealing (ACMAP [l]) using 50 iterations and (5) wavelet filter (3 levels of decomposition) [3]. Once both optical and S A R data were georeferenced, samples (- 300 relatively homogeneous parcels) were selected within the SPOT scene, thus allowing a better control of the sampling quality for the sensitivity study of filtering S A R data in the estimate of physical soil parameters. Angular dependence of the backscattering coefficient for HH polarization is given by [4]:
= 1w2.75.
Ohh

1
U

n o -.0.8
0.7

: :

0.6

I
0.5

0.4 0.3
0.2 Parcel

Box

Median

Gamma MAP

ACMAP

Wavelet

sin58 .

100.028.e.tan e

.(k.h). ~ i n ' . ~ B . h ~ . '

Figure 1: Mean value of correlation coefficients <r> between 0 ' from RADARSAT measurements and 0" values derived from the least-squares fit to the backscattering coefficient model (1). Error bars are for minima and maxima r values.
To quantify the impact of the speckle reduction technique on assessments of soil moisture and roughness, estimates of E and k.h derived from each of the 5 sets of angular filtered SAR images were compared with those from the parcel average method.

where 8 is the incidence angle, k the wave number, E the real part of the dielectric constant, h the rms height of the surface (cm) and A the wavelength (cm). The validity domain of this relationship corresponds to rms roughness (k.h)values within [0.3 ; 31 and incidence angles between

2138

: + Homogeneous parcels -;. ....-:. .....:. H e t e r o g e ~parcels s .

...........
...........
........... ...........

: I

7Oo1

.......................................................

- 1 ............................................
2

m :

!I

600

500

.......................

...........

-- .
-20

i
-18
-16

i
4
-12 -10

I
4

Box

Median

-8

-6

(b)

Gamma MAP

ACMAP

Wavelet

Estimsted d(dB)

Figure 2: Correlation between Gamma-MAP filtered 0" measurements (34 beam) and 0" estimated with the set of Gamm-MAPfiltered SAR data.
Whereas the mean relative deviations on E estimated values with all filtering methods are slightly constant (- 50 %) for homogeneas parcels, these deviations increase as the smoothing of the heterogeneous parcels decreases (Fig. 3a). same tendancies are observed for k.h estimated es (Fig. 3b), but with more accentuated fluctuations, The soil roughness is much more sensitive to the filtering approach. These large mean relative deviations on E and k.h estimates, especially for heterogeneous parcels, should 1 due to the good correlation be physically coefficients (r filtered 0 ' measurements and estimates derived from 412.
180...'..."."..

Figure 3: Mean relative deviation on ( a ) the soil dielectric constant and ( b ) the soil roughness estimated for each of the 5 filtering techniques and using average parcel method as reference.

CONCLUDING REMARKS
Results derived from this study stressed the importance of the choice of the S A R filtering technique for extracting the useful signal to be inverted for accessing information about soil dielectric constant and roughness. Depending on the speckle reduction method applied to raw intensity images, significant deviations were obtained on the estimated soil parameters in comparison with the commonly used parcel average method. The soil roughness is much more sensitive to the employed SAR filtering method. In fact, estimates of such a parameter using the parcel average method or non-adaptative filtering approach result from a too large smoothing of S A R intensity fluctuations within the parcel which seems physically significant for soil moisture and roughness retrieval.

..........................................

ACKNOWLEDGMENTS
We are very thankful to Pierre Vincent (VIASAT GCoTechnolgie Inc.) for the delivery of RADARSAT images.

REFERENCES
[ 11C. Oliver and S . Quegan, 1998, "Understanding synthetic

Box

Medbn

Gamma MAP

ACMAP

Wavelel

aperture radar images",Artech House Eds,Norwood (MA). 121 A. Lopes, E. N e w , R. Touzi and H. Laur, 1993, "Structure detection and statistical adaptive speckle filtering of S A R images", Int. J. Rem. Sens., vol. 14, pp.1735-1758. [3] S. Foucher, J.M. Boucher and G.B. BCniC, 1998, "Multiscale S A R filtering based on non-Gaussian assumptions", Proc. IGARSS98, Seattle (WA). [4] P.C. Dubois, J. van Zyl and T. Engman, 1995, "Measuring soil moisture with imaging radars", ZEEE Transac. on Geosc. & Rem. Sens., vol.33, pp.915-926.

9
~