Dr.Daniel Vidal-Madjar Co-Investigators

CETP M. Normand CEMAGREF

UVSQ O. Taconet CETP

10-12 avenue de l'Europe P. Dubois JPL/Pasadena

78140 Velizy S. Le Hégarat-Mascle CETP

France M. Zribi CETP

C. Loumagne CEMAGREF

C. Emblanch CEMAGREF

FINAL REPORT ON ORGEVAL' SIRC/XSAR 1994 EXPERIMENT

Test of Roughness and Moisture Algorithms Using Multiparameter Space Borne SAR and Application to Surface Hydrology

OBJECTIVES

Evaluate the usefulness of radar-derived parameters in surface hydrology

Demonstrate the usefulness of the squint mode in the case of bare soil observations

Compare various roughness/moisture algorithms in a real space imaging mode.

REPORT

I-Description of the Orgeval' SIRC/XSAR experiment in April 1994

1.1-The site

The Orgeval watershed, located 70 km East of Paris (France) is about 100 km2. It is an agricultural area with wheat, peas, corn crops and some forested areas. For the past 25 years, it has been extensively studied by the CEMAGREF (Centre National du Machinisme Agricole, du Genie Rural, des Eaux et des Forêts). The area is almost flat. The soil texture is uniform throughout the watershed with a clay and sand fraction of respectively 17% and 5%.

The experiment took place in April 94. Majority of the fields were bare soils, except for wheat (20 cm high). During the 5 days SARs passes, soil moisture remains high and constant (0.35cm3/cm3) over the watersheds and soil roughness and practices were the remaining factors of variability.

A global survey was done by elaborating crop map (70 fields of more of 1 ha) shown on Figure 1, photographies of soil surfaces over 50 bare fields. Ten fields (numbered 1 to 10) were intensively caracterized and were representative of the regional crops (4 of wheat, 2 of peas, 3 of corn, 1 ploughed field): soil moisture and density data (gravimetric and neutronic), heights profiles by 2 meter pin-profilers, penetrometer. A SPOT image (from June 1994 ) was used to assess the fields boundaries. The 70 fields were classified by roughness aspects (Table 1). The three classes are smooth surfaces, surfaces with intermediate roughness (or cloddy surfaces) and very rough surfaces.

On the ten selected fields, roughness parameters are derived from height profiles, half of them along the furrow direction and the other half in the direction perpendicular to the furrows. The 5 parameters are the RMS height of the surface, its correlation length (obtained from the parallel profiles), and three parameters describing the underlying periodic surface, the period, P, the width of the correlation function around the center frequency, L, and its intensity S. To compute the correlation length, l, an exponential correlation function is fitted through the data. The RMS height ranges between 0.6 cm and 3.77 cm. Measures are summed-up in Table 2.

1.2 SAR data

The SIR-C/XSAR system consists in an L-band and a C-band synthetic aperture radars (SAR) with quad polarization capabilities as well as a X-band VV polarization SAR. Table 3 lists all the data takes acquired over the Orgeval watershed over an incidence angle range of 44 to 57 degrees. Simultaneously several flights of the copolar scatterometer ERASME of the CETP (C and X bands ) were done in SAR site direction and over the tested region, keeping flights perpendicular and parallel to the rows directions over test fields. The ERASME incidence angles were from 25° to 50°.

Intercomparison of SIRC/XSAR's cross-sections was done over naturel targets,with fields seen on the same incidence angle and the same view azimuthal angle. The intercalibration of the 3 radars was within 2 dB at maximum (Figure 2).

SIGNIFICANT RESULTS

Polarimetric signatures in multiconfiguration conditions (3 frequencies, L, C, X, multi-incidence, and polarization ) were studied, following two approaches:

- a global analysis of availability of L and C band to discriminate (thanks to unsupervised classification algorithms) the polarimetric signatures of different land cover types, for crop maps elaboration in an early stage of plant growth

- the test of existing modelisations of backscattering over agricultural surfaces and the adequedation of surface soil description (moisture and roughness) derived from inversion of radar data by semi-empirical models.

II- Mono-band and multi-band classification results

2.1: Mono-band classification

We only report results corresponding to the SIR-C images (L and C bands), acquired in Spring, with incidence angles centred at 44°, 52°, 55°, and 57°. The data are available as multi-look complex images, with pixel spacing equal to 12.5m both in row and column directions. At the first incidence angle (44°), full quad-polarization image is available, but at the other incidence angles (52°, 55°, and 57°), only dual-polarization (HH and HV) images were acquired. Thus, we only consider HH and HV polarization power images for mono-band classification. Then , the polarimetric feature vector is:

where < > means averaged value over a 4 by 4 pixels window, chosen empirically as the best compromise between noise reduction and image feature preservation. The advantage of operating in the log domain is to give the image speckle the statistical characteristics of additive noise.

In Spring 1994, apart from winter wheat, which is still in a very early stage, there are mostly bare soils over the Orgeval site. For the supervised classification, we distinguish five different land cover types: forest, wheat, ploughed land, seedling and town, for which several training areas have been selected.

From the point of view of image information processing the supervised and unsupervised classification are fundamentally different. In the case of unsupervised classification no external information is used. Then classification results and therefore comparison results are closer to actual image information than in case of supervised classification where some ground truth information is introduced. For the unsupervised classification, cluster characteristics have been estimated by the Fuzzy C-Means (FCM) algorithm (Bezdek et al., 1984). The numbers of clusters are respectively equal to 4 for C band classification and to 6 for L band classification. The difference in the numbers of classes is due to the fact that the C band images are less contrasted than the L band ones.

In both cases, supervised and unsupervised classifications, blind classification and MAP with Markov Random Field (MRF) image model classification have been performed. In case of blind classification, no spatial information has been taken into account, and, in case of MRF image model, the label (or class) image has been modelled as a 8-neighbourhood Markov random field having Potts potential functions to favour homogeneous label configuration in a given neighbourhood (Geman and Geman, 1984).

Then identification rates of the main cover land types were computed. For each of the four land cover types: forest, wheat, ploughed land and seedling, Figure 3 (resp. 4) shows the C band (resp. L band) identification rates versus incidence angle, and the classification algorithm used: supervised (called "svs") or unsupervised (called "nsvs"), blind (subscript "0") or MAP (subscript "1"). The main results are:

a- taking into account neighbourhood information reduces error classification end leads to better identification rates;

b- L band performs better soil cover type discrimination than C band;

c- identification rates in C band increase with incidence angle (between 44° and 57°);

d- ploughed land identification in L band decreases with incidence angle.

2.2- Complementary and redundancy evaluation

Then, complementary and redundancy between L and C bands were studied in a quantitative way. We applied the measurement of redundancy defined in Shannon (1963) information theory. Let us consider two images (sets of messages) X1 and X2 having values respectively in N1, N1= and N2, N2 = .The mutual information redundancy is:

where p(x,y) is the joint probability of x and y, and p(x / y) is the conditional probability of x knowing y.

Moreover, we compute redundancy at class level (using classification results), because it presents the advantage of considering a higher level of information than radiometric information level, for example the spatial information may be taken into account if redundancy is computed between MAP classification using MRF image model. However, we have to keep in mind that the level of information, at which the comparison is performed, is dependent on the chosen image model and on the classification algorithm used.

Figure 5 shows the redundancy rate (mutual information normalised by the image entropy) between L and C band unsupervised classification results, versus incidence angle. We found that redundancy increases when incidence angle increases between 44° and 57°. However at incidence angle equal to 57°, redundancy between L and C band is only about 30%, which suggest great interest of data fusion.

2.3-Data fusion and multi-band classification

At last, data fusion between L and C band was performed to improve identification rates. The aim of data fusion is to use redundancy between images to reduce classification errors and to use complementary to put in evidence new classes. Unsupervised multi-source classification algorithm used is based on Dempster-Shafer evidence theory. Figure 6 shows the identification rates obtained after data fusion, and improvement in identification of land cover types clearly appears.

The conclusion is that multi-band L and C can be successfully used to discriminate the different land cover types. Similar results were obtained with AirSAR data for discrimination of the different culture types in Summer 1991.

III-Backscattering modelling and retrieval of surface soil parameters by inversion of radar data

3.1- Comparison between simulated and measured radar data

The complementary use of ERASME scatterometer and SIRC/XSAR's allows to study in copolarization the variation of radar cross sections with incidence angles from 25 to 57 degrees. Over the 10 tests fields, the surface roughness parameters were deduced from the height profiles using the Shin and Kong' 1984 quasi periodic description of agricultural soils as proposed by Rakotoarivony et al (1996). Small spatial scale (clods) and large ones (rows) parameters are summed in Table 2 and localised in ks/kl space (k the wave number, s the r.m.s. heights, l the length of correlation at small scale) on Figure 7. In L band, all test fields are within IEM validity range. In C band, only smooth ones (peas) are within. Most other ploughed fields are out of range. Field W3 surface characteristics are taken for eroded smooth soil, W2 for smooth ones, and C9 for ploughed one remaining in border of IEM limits.

Based on experimental observations and analysis of theoretical models, semi-empirical models were developed by Oh et al (1994) and Dubois et al (1995) for the backscattering coefficients, in terms of surface dielectric constant and a surface roughness parameter (RMS heights). Validity limits are respectively:

-for Oh model, 0.1<ks<6 and 2.5<kl<20

-for Dubois model, ks<2.5 and 30<I<60 degrees.

Results of comparison between the three models simulations and ERASME, SIRC/XSAR's data are shown on Figures 8, 9 and 10. They arethe following ones.

a- In L band there is a good agreement between radar data and IEM outputs for smooth and cloddy surfaces within 2dB. The agreement remains good between Dubois and IEM for incidence angles smaller than 45 degrees. The decrease of radar cross sections with incidence angles is also well reproduced by the Oh algorithm but with a constant underestimation of 5dB.

b- In C band IEM simulations remain good for the smooth fields between 25 to 40 degrees. For larger incidence angles (40 to 57 degrees), there is a systematic overestimation of the model. IEM underestimates the slope for ploughed surfaces resulting in a discrepancy that can reach 10dB. Oh and Dubois outputs are very close together and reproduce in a right way the slope decrease of radar data with incidence angles but with an underestimation of 5dB.

c- In X band the Oh model remains the only able to simulate the microwave measurements. A slight underestimation of 2dB is noted for rough surfaces.

Even though these semi-models were derived using data sets different from the Orgeval considered one, this analysis gives confidence that they could be applied to other data sets and precises their adequacies and limits. It is a necessary step before using them to define the inversion of radar measurements to specify the soil description.

3.2- Evaluation of the inversion algorithm of Dubois et al (1995)

3.2.1- Discrimination between the three roughness groups using L and C bands data

The data take at 44 degree(14 April) is the only one in. full polarization, giving HH and VV intensities. We used the inversion algorithm (Dubois et. al., 95) to estimate the RMS height and the dielectric constant of the surfaces of the bare fields. The soil moisture content was then computed from the dielectric constant using the Hallikainen curves (Hallikainen et. al., 1989) for a soil texture of 17% clay and 5% sand. The results are summarized in Figure 11.

Overall, we notice that L and C band data provide consistent information about the roughness of the surfaces. Here the considered RMS height is the one representative of the small scale roughness (clods) computed on the height profiles collected along the furrows. At L band the three roughness groups can easily be discriminated by the estimates of the RMS height. In C band when comparing the surfaces of intermediate roughness to the rough surfaces, the two groups overlap while they are almost dissociated at L band.

For six fields fully characterized using the profilometer, the measured RMS height is compared to the estimated RMS height in Figure12. The resulting RMS error is 0.45 cm at L band and 1.26 cm at C band. At L band, the RMS error is dominated by the rougher data points. At C band, the RMS error is dominated by the error on the RMS height estimate for Field 8.

The estimates of RMS height for Field 8 are very different at L-band than at C-band. The L-band estimation is close to the measured RMS height while the C-band estimate is much higher. Field 8 is a field with a strong periodicity (around 40 cm) perpendicular to the radar look direction. In Table 4, we have indicated the Bragg modes for Lband and Cband. According to the small perturbation model principles, the parts of the surface spectrum corresponding to these Bragg modes are contributing the most to the scattering. The Cband Bragg modes are accessing the part of the spectrum where the periodicity peak is, resulting in an enhanced scattering. The L band Bragg modes are falling outside this hump resulting in an estimation of the RMS height close to the one estimated using the parallel spectrum.

3.2.2- Comparison between L band and C band roughness estimations

In Figure13, the C band estimates for the RMS heights are plotted versus the corresponding L band estimates. There is a strong correlation between the estimated RMS height at C band and at L band with two definite regimes. Some points fall close to the 1:1 line while for others, the estimated roughness is much higher at C band than at L band. A possible explanation to these two regimes is proposed when looking at the characteristics of the surfaces. For the first regime the surface is rather rough and was worked recently, as for the second regime the surfaces have not been worked recently and are eroded. The initial weathering of a surface was observed to fill up the lower parts of the surface with fine material resulting in a smoother surface with a different structure, characterized by the presence of clods emerging from a very smooth surface. The differences between these two types of surfaces is very apparent but not easy to quantify.

In an effort to do so, we have devised the following measure based on the spectra of the surfaces. From the surface spectrum acquired in the direction parallel to the row direction, we compute a "L band normalized" surface spectrum by dividing the surface spectrum by its maximum value taken in a narrow interval around the L band first Bragg mode. Then, we compute the average of this new spectrum between the L band first Bragg mode and the C band first Bragg mode. It quantifies the amount of energy in the spectrum visible only by C band relative to the part of the spectrum contributing the most to the L band response. This new parameter will be called the 'C band over L band scattering indicator' (COLSI).

The data collection of height profiles allows the analysis of eight measured spectra: four corresponding to the rough surfaces and four corresponding to the eroded surfaces (one bare and three below wheat cover). In Figure 14, we have plotted the "L band normalized" power spectra between the L band and the C band first Bragg modes. Spectrum slope seems to be flatter for the recently-worked surface. In Figure 15, we have plotted the COLSI as a function of surface type. It leads to COLSI index larger for the eroded surfaces than for the recently worked rough surfaces.

The previous validations for the empirical model of Dubois et al (1995) have been done essentially at L band over all types of surfaces including (non) and eroded ones and it appears that the estimations of RMS heights using L band data are satisfactory with an estimated RMS error of about 0.4 cm. The sensitivity of the L-band scattering to the surface erosion level is weak. We believe that the robustness of the L-band estimation of RMS height comes from the relative stability of the lower frequencies of the spectrum (L-band access the lower frequencies) with erosion but this certainly needs to be further investigated. However, the validity of C band RMS height estimations depend on the surface type, eroded or non-eroded. This seems to indicate that at C band, the surface scattering cannot be characterized only by the knowledge of the RMS height. The COLSI was shown to be able to provide some of the missing information by discriminating between the two surface types. COLSI is dependent on the power spectrum of the surface and is certainly related to the mean slope of the spectrum. The other classical surface parameters such as mean RMS slope or correlation length could be compared to COLSI. However, we feel that the differences between the weathered and non-weathered surfaces are more visible in the higher frequencies of the spectrum while theseen other parameters are usually estimated using the low frequency end of the spectrum.

3.2.3-Estimation of the dielectric constant and soil moisture

The inversion algorithm used in the preceding section produces an estimate of the roughness and of the dielectric constant given the HH and VV returns. Unfortunately, the other data takes, because of their high incidence angle could not be acquired in copolarised intensities. The HH return and the HV return are the only available signals. Then we assume that the RMS height is constant throughout the five days and can be computed using the data acquired on April, 14. From the value of the RMS height and the HH signal, we compute the dielectric constant and the soil moisture content over the studied fields. The trend of the average soil moisture content over the fields is plotted in Figure 16. In Figure 16, we can see that the estimated soil moisture follows well the slight drying trend of the measured soil moisture. Unfortunately, on April 17, 1994, marked 2 on Figure 16, no measurement was made.

REFERENCES

Bezdek, J.C., Ehrlich, R., and Full, W. (1984), "FCM: the fuzzy c-means clustering algorithm", Computer & Geosciences, vol.10, no.2-3, pp. 191-203, 1984.

Dubois, P., Van Zyl, J. J. and T. Engman (1995) "Measuring soil moisture wiyth imaging radars", IEEE Transaction on Geoscience and Remote sensing, Vol.GE-33, 915-926, 1995.

Geman, S., and Geman, D. (1984), "Stochastic relaxation, Gibbs distribution and Bayesian restoration of images", IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.6, no.6, pp.721-741, Nov. 1984.

Hallikainen, M. T., F.T.Ulaby, M.C.Dobson, M.A.El-Rays and L.K.Wu (1985), "Microwave Dielectric Behaviour of Wet soil - Part I: Empirical Models and Experimental Observations": IEEE Transaction on Geoscience and Remote sensing, Vol.GE-23,n°1,January 1985

Mascle, S., Bloch, I., and Vidal-Madjar, D. (1995), "Unsupervised multi-source remote sensing classification using Dempster-Shafer evidence theory", in Proceedings of Europto Conference on Synthetic Aperture Radar and Passive Microwave Sensing, in Paris, France, on September 25-28 1995, SPIE vol.2584, pp. 200-211.

Oh, Y., K.Sarabandi and F.T. Ulaby (1994), "An inversion algorithm for retrieving soil moisture and surface roughness from polarimetric radar observations", IEEE Transactions on Geoscience and Remote Sensing, 1582-1584,1994.

Rakotoarivony, L., Taconet, O., Vidal-Madjar, D., and Benallegue,M. (1996), "Radar backscattering over agricultural bare soils," J. Electro. Waves Applic., Vol. 10, No. 2, 187-209, 1996.

Shafer, G. (1976), "A Mathematical Theory of Evidence", Princeton University Press, 1976.

Shannon, C.E., and Weawer, W. (1963), "The mathematical theory of communication", University of Illinois Press, Urbana, 1963.

Shin, R. T. and J. A. Kong (1984) "Scattering of Electromagnetic Waves from a Randomly Perturbed Quasi-periodic Surface", J. App. Phys., 56 (1), July 1984.

MASCLE, S., VIDAL-MADJAR, D., ZRIBI, M., and TACONET, O., "Comparison between L and C bands SIR-C polarimetric data versus incidence angle", in Proceedings of International Symposium: "Retrieval of bio- and geophysical parameters from SAR data for land applications", in Toulouse, France, on October 10-13 1995, pp.273-281.

TACONET,O., D. VIDAL-MADJAR, C. KING, Y. LE BISSONNAIS, M. ZRIBI, S. MASCLE, C. LOUMAGNE and M. NORMAND, Soil backscattering behaviour with roughness from combination of SIRC/ XSAR imagery and airborne scatterometer data (ERASME and RENE), IGARSS'96, Burnham Yates Conference Center, Lincoln, Nebraska, USA, 27-31 May 1996.

LE HEGARAT-MASCLE, S., BLOCH, I., and VIDAL-MADJAR, D., "Application of Dempster-Shafer Evidence Theory to Unsupervised Classification in Multisource Remote Sensing", in IEEE Transactions on Geoscience and Remote Sensing, accepted 11 December 1996.

LE HEGARAT-MASCLE, S., VIDAL-MADJAR, D., TACONET, O., and ZRIBI, M., "Application of Shannon Information Theory to a Comparison Between L and C Bands SIR-C Polarimetric Data Versus Incidence Angle", in Remote Sensing of Environment, accepted 6 August 1996.

ZRIBI, M., O. TACONET, D. VIDAL-MADJAR, S. MASCLE, C. LOUMAGNE and M. NORMAND, Backscattering response with direction angle relative with soil practices from combination of SIRC/ XSAR imagery and ERASME airborne scatterometer data (Orgeval'94), in preparation for Remote Sens. of Environment, 1997.

TABLE 1: Classification of the Orgeval fields by soil covertures and roughnesses.

CLASS	FIELDS	DIRECTION TOWARDS ROWS FOR SIRC/XSAR
WHEAT	2,3,71,78 4,5,62 67,83,85,68	Perpendicular Parallel Between
ERODED SMOOTH BARE SOILS	10,12 40 23,18,31,42	Perpendicular Parallel Between
SMOOTH BARE SOILS	35 11,46 16,47,53	Perpendicular Parallel Between
PLOUGHED FIELDS	6, 8 1,9,7,49	Perpendicular Parallel

TABLE 2: Roughness parameters

Field	s	l	S	L	P	ks (L)	kl (L)	ks (C)	kl (C)	ks (X)	kl (X)
C1	2.66	6.1	0	0	0	0.70	1.6	2.95	6.77	5.28	12.1
W2	0.70	5.97	0.97	34.69	99.58	0.18	1.57	0.78	6.63	1.40	11.9
W3	0.55	12.68	0.73	143.38	197.31	0.14	3.33	0.61	14.08	1.09	25.2
W4	0.92	2.51	0.83	42.37	104.56	0.24	0.66	1.02	2.79	1.82	4.8
L6	3.77	7.35	0.94	128.87	94.35	0.99	1.93	4.19	8.16	7.5	14.6
C7	2.49	8.1	1.73	147	45.2	0.65	2.12	2.76	8.99	4.94	16.9
P8	1.31	3.52	1.44	93.67	43.0	0.34	0.92	1.45	3.9	2.59	6.9
C9	0.97	8.07	1.63	115.26	40.22	0.25	2.12	1.07	8.96	1.91	16.2
P10	0.62	5.66	0.3	172.08	120.68	0.16	1.48	0.68	6.28	1.21	11.2

W:Wheat, C: Corn, P: Peas, L: Ploughed field

TABLE 3: SIR-C data takes

Data take	Incidence Angle	Mode	date
11332	26°	HH,VV,HV	4/12
11334	34°	HH,VV,HV	4/13
11914	44°	HH,VV,HV	4/14
	48°	HH, HV	4/15
11919	52°	HH, HV	4/16
11560	55°	HH, HV	4/17
12101	57°	HH, HV	4/18

The first two data takes unfortunately did not cover the study area. The data take from 04/15 could not be used in the study because of some calibration problem.

TABLE 4: Bragg modes for L and C bands at 44 degrees

Bragg Mode Number	1	2	3	4	5	6	7	8	9	10	11	12
K (C band) cm	4.0	8.1	12.1	16.1	20.1	24.2	28.2	32.2	36.3	40.3	44.3	48.4
K (L band) cm	17.3	34.5	51.8	69.1	86.3	103.	120.	138.	155.	172.	189.	207.

FIGURE 12