ERS SAR time series analysis for maize monitoring using experimental and modeling approaches

Abstract

The study aimed at documenting the effective ERS-1 capabilities to monitor maize growth using field and SAR measurements over a large area (6000 km²). During the summer 1995, a first ground campaign was carried out to determine the row orientation and the relief of 555 fields. Secondly, bio- and geophysical parameters such as fresh and dry biomass, plant height and soil moisture were measured on 75 fields simultaneously to SAR image acquisitions. In total, 12 ERS-1 products were acquired. The results show that no specific maize crop temporal backscattering signature can be determined and that the soil moisture still influences the backscattering when the crop has reached is maximum height. However, the temporal evolution of the averaged field variation coefficient revealed the effect of the vegetation. A stratification approach was thus used to account for the soil moisture and to determine the crop parameters effect on the signal. A correlation coefficient of 0.9 between the s° (dB) and the dry matter quantity per ha was obtained on fields larger than 2 ha. Finally, the row orientation was show to have an influence on the attenuation by the canopy. These findings lead to the development of a water cloud model formulation dedicated to crop with a row-structured canopy.
Keywords: ERS-1, SAR, maize monitoring, temporal signature, modeling

Introduction

SAR remote sensing could play a major role in agriculture monitoring. However, exploitable results in an operational framework are still expected particularly for early determination of crop acreages and yield forecasting. This ESA Pilot Project aims at documenting the effective ERS-1 capabilities to monitor crop growth cycle using theoretical modeling and a comprehensive data set of field and SAR measurements. A single crop was chosen to allow intensive measurements and the maize case was selected because it is considered as one of the most transparent crop. This paper briefly presents the results of this two year study.

Based on a field campaign focused on a single crop, the study objectives are to document the various factors influencing the signal, i.e. soil moisture, vegetation cover and field size, and to assess the signal sensitivity to crop parameters. Furthermore, the row-structure effect on the signal is also investigated using modeling and empirical approaches.

Study area

The study area was selected as large as possible to be representative of the maize crop in Belgium. It covers a total surface area of 6000 km² (see fig. 1). The soil is constituted of a Pleistocene loam deposit with a very stable textural composition (80 % of loam, 10 to 15 % of clay and 5 % of silt). The relief is slightly undulating and the mean slope of the fields is about 2 or 3 per cent. These characteristics allow us to avoid major interference related to soil type and topography.

Figure 1: Localization of the 555 test fields and of the two image sets over the study area.

Ground truth measurements

Ground truth data have been collected in a two-step field campaign. During the first one - at the beginning of the growing season - 555 maize fields were identified and localized (see fig. 1). For each, we recorded the row direction, the crop type of the surroundings, a rough scheme of the relief and the mean slope of the field. According to this last parameter, the fields were classified into three distinct categories (slope <1%: no relief; slope = 2 or 3 %: light relief; slope >4%: high relief)

A sub-sample of 76 maize fields constitutes the data set of the fully monitored maize fields for the second ground campaign. These were systematically visited at each of the satellite overpasses between mid-June and mid-October to proceed to biophysical parameters (leaf stage, plant height, wet biomass and plant moisture content) and gravimetric soil moisture measurements.

Daily meteorological parameters between 1st April 1995 to 31st October 1995 were also collected thanks to the national meteorological network. Temperatures, rainfall, sunshine and solar radiation data are respectively available for 11, 19, 4 and 1 meteorological stations.

ERS-1 SAR data

The whole study area is covered by two contiguous and descending SAR ERS-1 set of images. The revisiting period is 35 days but the common area of the two sets is overflown twice the period. Twelve images - in fact, six of each set - have been acquired between April and October 95. Figure 1 shows the localization of the two contiguous sets over the study area.

The data provided in a PRI format have been calibrated on a pixel basis using calculations stemmed from the procedure of H. Laur (1992). The two sets of images were separately geo-referenced in order to facilitate the localization of the fields. This was made thanks to the SPOT mosaic and partly to a set of airborne pictures acquired in May 95. The fields boundaries were drawn on each multi-temporal SAR image and used to extract for each field and each date :

• the backscattering coefficient in m²/m² : ;
• the standard deviation of the field pixels : ;
• the backscattering coefficient in dB : .

where:

• N is the number of pixels for one field;
• 

is the backscattering coefficient in (m²/m²) for pixel i.

Results

Temporal signature of maize fields

The temporal backscattering signature is shown on figure 2. Each point represents the mean s° (dB) for the 150 fields located in the overlap area of the two image sets and the vertical bars, the standard deviation of the sample. This curve has a saw-tooth profile which can not been explained by the evolution of the crop biophysical parameters also represented on the graph. As already illustrated by Clevers et al. (1996) for other crops, the mean backscattering coefficient time profiles for maize will not present any standard pattern comparable fromyear to year.

Figure 2: Temporal evolution of the averaged s°(dB) obtained from the 150 fields located in the overlap area of the two image sets. Vertical bars represent the standard deviation of the samples. Main crop phases and the temporal evolution of the biomass are also illustrated. At the end of the season, the mean biomass is calculated on fields which have not been harvested.

As documented in the litterature (Engman, 1991), the behavior of the ERS-1 signal is found to be related to the soil moisture variability. Figure 3 presents on the same graph the temporal signature of the signal and the evolution of the precipitation index defined as the sum of the precipitations of the six days before the image acquisition. Peaks and drops in the radar backscatter curve corresponds to periods of dryness and wetness. Thus, the maize canopy must be considered as partly transparent to the microwaves even at fully developed stages.

Figure 3: Comparison of the temporal evolution of the mean backscattering coefficient and of the mean precipitation index.

Sensitivity to soil moisture

To confirm the soil moisture effect during the vegetation phase of the crop,. the backscattering coefficient has been plotted against the gravimetric soil moisture for 67 fields on figure 4. The test-sample has been set up with the two following constraints:

• first, a sufficient range of humidity has to be covered which imposes to work on measurements collected at different dates;
• secondly, the soil moisture has to be the main variable. Consequently, the crop status at the selected dates must be similar in order to avoid important biomass variations.

Moreover, no field repetition is included to keep the sample independancy.

Figure 4: s° (dB) against gravimetric soil moisture for the 67 selected fields. The correlation coefficient r is equal to 0.63

The correlation coefficient between the two variables is greater than 0.6. Assuming a linear relationship, it means that 40% of the sample variability is explained by the soil moisture. Nevertheless, the remaining variability is still important.

At this stage, the sample contains some fields of very small size. According to the theory (ESA, 1995), this could lead to a significant uncertainty due to speckle noise. In figures 5 and 6, we plotted two sub-samples according to the size of the fields. The limit has been fixed at 192 pixels or 3 ha for a PRI image.

Figure 5: s° (dB) against gravimetric soil moisture for the 25 fields with size smaller than 3 ha. r1 = 0.42

Figure 6: s° (dB) against gravimetric soil moisture for the 43 fields with size higher than 3 ha. r2 = 0.76

The correlation coefficient of the first sub-sample is almost two times smaller than the one of the second population. A test of equality of the correlation coefficients was made for the two populations. The null hypothesis (r1=r2) is rejected at a level of 0.05 and the difference can be considered as highly significant.

Vegetation effect

In spite of the influence of soil moisture, the temporal signature of the averaged variation coefficient shows a pattern compatible with a vegetation effect (see fig. 7). The mean variation coefficient has been computed for 150 fields as follows :

where :

• M is the number of fields, i.e. 150;
• is the standard deviation of the pixels for the field j;
• 

is the backscattering coefficient (m²/m²) for the field j.

According to statistical properties of radar images, the standard deviation of a target without texture is proportional to the mean and the variation coefficient is a constant equal to 0.58 (=1/ûN where N is the number of look of the radar image and is equal to 3 in this case). For a textured target, the variation coefficient increases with its heterogeneity (Laur, 1989). The mean variation coefficient is higher than 0.58 during all the season but presents a strong decrease corresponding to the canopy closure between day 180 and day 200. This characteristic pattern is not visible on temporal signatures of the variation coefficient for single fields.

Figure 7: Temporal evolution of the variation coefficient averaged for 150 fields.

To integrate the crop biophysical parameter influence on the signal, the soil contribution had to be accounted for. This has been done using a stratification approach for the setting-up of test samples. Fields were distributed among different classes on the basis of their soil moisture. As in the previous section, measurements realized on the same field at different dates were avoided and smaller fields (inferior to 2 ha) were rejected.

On figures 8 and 9, the backscattering coefficient has been plotted respectively against the dry matter quantity per hectare (dmh) and the dry matter quantity per unit of volume (dmv) for the only class containing sufficient data. The 16 selected fields have a soil moisture between 10 and 14 % and comes from three SAR scenes. In the first case, the relationship between the two variables seems to correspond to a negative exponential law. For the second canopy parameter, the vegetation height is accounted for in the denominator and the relationship between the two variable is approximately linear. For this graph, a correlation coefficient of 0.90 was found.

Figure 8: Backscattering coefficient (dB) as a function of the quantity of dry matter per ha (dmh) for the 16 fields of the sample.

Figure 9: Backscattering coefficient (dB) as a function of the dry matter quantity per unit of volume (dmv) for the 16 fields of the sample. r = 0.9.

Further tests were made to assess the effect of the row direction on the signal. Indeed, the maize crop has a row-structured canopy clearly visible during the early phenological stages of the crop. First, the previous sample was enlarged to account for fields with soil moisture going from 7 % to 14% and the 28 fields finally retained were divided into two groups depending on the row orientation.

For fields with a relative row orientation between 0 and 45°, figure 10 shows that the relationship between the backscattering coefficient and the dry matter quantity hectare is worse than on figure 8 while it is not affected for fields having a relative row orientation between 45 and 90° (see fig. 11). This could be related to the fact that the number of rows to be penetrated by the radiation increases with the view angle.

Figure 10: Backscattering coefficient in dB as a function of the quantity of dry matter per ha (dmh) for 15 fields with soil moisture between 7 and 14 % and with a row-direction between 0 and 45°.

Figure 11: Backscattering coefficient in dB as a function of the quantity of dry matter per ha (dmh) for 13 fields with soil moisture between 7 and 14 % and with a row-direction between 45° and 90°.

The canopy structure effect is also confirmed on figure 12 where the s0 (dB) has been plotted against the radar look direction , that is to say the azimuth angle f between the ground projection of the antenna beam, for 33 fields with soil moisture and biomass considered as similar.

Figure 12: Plot of the backscattering coefficient against the look direction f for 33 fields located near each other. The backscattering coefficients were extracted on the SAR scene of the 13/06/95, after the emergence of the crop; the canopy height is approximately 40 cm. A look direction of 0° signifies that the beam of the incident angle is parallel to the row direction.

Modeling approach

Considering these results, the water-cloud model (Ulaby et al., 1986) has been extended to crops which present a row structured canopy at early growth stages such asmaize and sugar beet.

As in the original formulation, the total backscattered power is computed as the incoherent sum of both the underlaying ground and the vegetation contribution where:

The quantity repressent the backscattering over the bare soil. However, in the new formulation, the vegetation backscattering coefficient is no longer integrated on the whole height of the canopy.

On figure 13, we see that the canopy is considered as a succession of adjacent rows of height h, row-spacing D, and row-width d. Other parameters are the incidence angle q and the angle Y between the row-direction and the SAR view angle. As shown on figure 14 for the particular case Y = 90°, the triplet of parameters {D, d, htanq} allows us to discriminate between growth stages.

Figure 13: Row-structure of the canopy.

Figure 14: Scenario associated to the growth stages to be considered in the modeling. Radiation are orthogonal to the row direction, i.e. y = 90°

Because of the canopy structure, the radar cross-section (rcs) depends on the position x along the axis with period D. The simple shape of the canopy implies that the total rcs is the incoherent sum of four terms , each one being the solution over a spatial subinterval . We thus write the backscattering coefficient as follows:

,

with:

,

where the calculation of the backscattering coefficient follows the original formulation by Ulaby et al (1986), but with the bounds of integration over the (Oz) axis changing accordingly to the intervals . The bounds of the four domains depend on growth stages, incidence angle, crop characteristics, i.e. large or small row-width in comparison to either the row-spacing or the quantity , and on the amount of rows which interacts with the radar signal. The details and the whole set of solutions has been described by Baltazart et al. (1996).

First simulations have shown the negative trend of the two-way attenuation through the vegetation layer with increasing y angle from 0 to 90 degrees, and a positive trend of about the same magnitude for the backscattering by the vegetation alone, i.e. . However, the trend of the total backscattering coefficient appears to be mostly driven by the backscattering sensitivity over bare soil. Backscattering sensitivity to look direction over bare soil varies the same way as with a magnitude depending on the tillage and drainage patterns and seed bed preparations. The vegetation cover gradually conceals most of this sensitivity. This is illustrated over sugar beet crops by the following simulation on figure 15.

Figure 15: Contribution of the different terms to the backscattering coefficient. Extinction coefficient = 2.5 Np / m, scattering coefficient = 2.2 m-1 , canopy height h = 0.3 m, row-spacing D = 0.4 m, row-widths d=0.2 m, frequency =5.3Ghz, incidence angle q = 23 degrees and VV polarisation. s0 of the bare soil is deduced from Müller et al.(1993).

Conclusions

The aim of this project was to assess the capabilities of the ERS-1 SAR data to retrieve quantitative biophysical and geophysical variables in the perspective of crop monitoring. The methodology was based on the complementary contributions from theoretical and empirical approaches.

The theoretical approach focused on the understanding of the backscattering by row structured crop canopies at early growth stages (e.g. maize, sugar beet, vineyard and sorghum). The research leads to the development of a semi-empirical model accounting for the structure and orientation of the canopy rows and which could be used for the early prediction of crop yield. The empirical study was based on a very intensive ground campaign set up during summer 95 over 550 maize fields. The collected data were put together in a relational data-base with meteorological data and ERS-1 backscattering coefficients extracted from 12 ERS-1 SAR images acquired over the whole season.

Due to the structure of its canopy, maize is generally be considered as one of the most transparent crops to microwaves. Soil properties under maize canopy influence the signal backscattering all over the growth season. As a consequence, the mean backscattering coefficient time profiles for maize crops do not present any standard pattern comparable year to year. This study confirms the sensitivity of the ERS-1 SAR signal to soil moisture. For fields larger than 3 ha at full vegetation stage, i.e. canopy at its maximum height, acorrelation coefficient of 0.76 was found.

However, the temporal evolution of the averaged variation coefficient of the fields revealed some effect of the vegetation. A stratification approach was used to account for the soil moisture and to assess the effect of the crop parameters on the ERS-1 signal. A correlation coefficient of 0.9 between the s° (dB) and the dry matter quantity per ha was obtained on fields larger than 2 ha. This minimum field size requirement was imposed for reducing the speckle effect on the estimation uncertainty.

Finally, further work focused on the study of the row direction effect on the ERS-1 SAR signal. We found that during the early growth stage of the crop, backscattering coefficients extracted from large flat fields with homogeneous soil moisture conditions can vary of approximately 2 or 3 dB in relation with the canopy row orientation.

Acknowledgments

The authors would like to thank to the Belgian Federal Office for Scientific, Technical and Cultural affairs for his financial support and ESA organization for providing SAR images in the framework of this pilot project.

References

Baltazart, V., Auquière, E., Guissard, A., Defourny, P., 1996 (submitted):
On The Extension Of The Water Cloud Model To A Row-Structured Canopy, in IEEE on Geoscience and Remote Sensing.

Clevers, J., van Leeuwen, H., Hoekman, D., Nieuwenhuis, G., Kramer, H., Vissers, M., 1996:
Application of JERS-1 data for agriculture, forestry and land use, NRSP report 95-33 of the BCRS project 3.2/TO-06, Netherlands Remote Sensing Board, Delft, 95 p.

Engman, E.T., 1991:
Applications of Microwave Remote Sensing Of Soil Moisture for Water Resources and Agriculture, in Remote Sensing of Environment, Vol. 35, n° 12, pp. 213-226.

ESA SP-1185, 1995:
Satellite Radar in Agriculture, Experience with ERS-1, ESA Publications Division, Noordwijk, 71 p.

Laur, H., 1989:
Analyse d'images radar en télédétection: discriminateurs radiométriques et texturaux, Thèse, Université Paul Sabatier, Toulouse.

Laur, H., 1992:
Derivation of backscattering coefficient in ERS-1 SAR PRI products, ESA-ESRIN, Issue 1, Rev. 0, 17, 17 p.

Müller, U., Löcherbach, T., Förstner, W., Kühbauch, W., 1993:
Suitability of ERS-1 SAR PRI-data for multitemporal agricultural land-use mapping, in Proc. ERS-1 Symposium, ESA SP-361, Noordwijk, pp. 69-74.

Ulaby, F.T., Moore, R.k., Fung, A.K., 1986:
Microwave Remote Sensing, Vol. III, Artech House, pp. 1065-2162.

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