METHODOLOGICAL ADVANCEMENTS IN USING ERS SAR DATA FOR CROP AREA ESTIMATION.

ABSTRACT

The paper reports the results of methodological research into the use of multi-temporal ERS SAR data for crop area estimation. This work has been carried out as part of the pilot project on the use of ERS SAR data for rapid area estimates of agricultural crops which was initiated by the DG VI of the European Commission with the technical and scientific support from the MARS Agricultural Information Systems unit of the Joint Research Centre. In the paper, an innovative methodological framework is presented to generate multi-temporal ERS image composites for optimal crop classification. The use of this method for early season and full season time series is presented and discussed. We show that soil moisture and surface roughness induced variation in backscattering signatures needs careful consideration in a crop classification approach. The use of calibrated geocoded and speckle filtered ERS image data has lead to a classification accuracy, which for the given sites and crops, can be considered superior than those obtained with SPOT optical/IR data. It is shown that the further integration of ERS-derived crop area estimates in the MARS operational activities is feasible and recommendable, under the condition that certain operational constraints are met.

Keywords: Agriculture, crop classification, constraints, MARS project

INTRODUCTION

The use of ERS SAR data for early season crop area estimation in the Monitoring Agriculture with Remote Sensing (MARS) program has reached a pre-operational status with the launching of an effort to produce area estimates in winter and spring for 60 sites in Western Europe. This effort is financed directly by the Directorate General VI (Agriculture) of the European Commission, and technically and scientifically supervised by the Agricultural Information Systems Unit of the MARS project of the Institute for Space Applications at the Joint Research Centre.

The 1997 activity, which is carried out by a consortium lead by NRSC Ltd (UK) and including SYNOPTICS, Integrated Remote Sensing and GIS Applications BV , follows a dedicated research effort in 1995 (SYNOPTICS, 1996a) and an experimental evaluation phase in 1996 (GAF et al, 1997). During the 1995 and 1996 projects, a number of critical items and constraints on the use of SAR data for crop area estimation have been addressed. This has lead to (1) the definition of the main requirements for SAR post-processing, (2) the development of a novel method for interpretation of SAR time series, and (3) the configuration of an interpretation environment in which a range of ancillary products are introduced in support of classification of temporal SAR signatures. The first of these items is further detailed in Sowter et al, 1997, and the latter in Kidd et al, 1997. This paper is concerned mainly with the description of the interpretation method.

The paper is structured as follows: in the following section, the methodology is described. The performance is illustrated with two examples for the Great Driffield (UK) and Bernburg (Germany) site. In the discussion, we address the accuracy in comparison with SPOT data for the same site, and illustrate some other interesting applications which are possible within the interpretation chain. We conclude with some remarks on expected further developments.

METHODOLOGY

The Activity B of the MARS project is focused on deriving crop area statistics for 60 sites in Western Europe. The operational of this exercise is based on the analysis of optical imagery during the growing season. For each site, between 1 and 4 optical images are acquired for producing area coverages for all sites. The use of optical imagery allows production of the first reliable area estimates starting at the end of April after the first images have been processed. The use of optical images requires an appreciable reflection signal (hence, sufficient illumination by the Sun) and vegetation development (differentiation between bare and vegetated fields). Severe constraints due to cloudiness and haze are well known.

It has been recognised at an early stage that the use of SAR might advance the earliest date of crop area generation because microwave backscattering that underlies SAR image formation is sensitive to surface roughness change as well. Thus, in principle, the separation of soil surfaces by roughness type could lead to an early indication of generic crop classes, such as grass land, winter crops (smooth seedbeds at the onset of winter) and spring/summer crops (rough fields). Furthermore, in a multi-temporal approach, one should be able to follow transitions in surface roughness, which are indicative of soil tillage, and, while crop-specific, useful in further specification of crop classes. The potential of using autumn and spring SAR imagery for crop area estimation was reported previously (Lemoine et al, 1995, Nezry et al, 1996).

Because backscattering values are geo-referenced and calibrated it is possible to carry out SAR image interpretation in a quantitative manner, as long as a number of ancillary data are available. Information content in single SAR images is often limited, at least for agricultural scenes, and especially in the autumn to spring period. Therefore, a multi-temporal approach is usually adopted. A first advantage of such an approach is the possibility to reduced speckle, but also to follow transitions over time with calibrated backscattering signatures. In most studies, multi-temporal time series are combined in a composite and then classified, usually after some clustering has been performed. The success of such a classification is dependent on the parameter settings of the clustering algorithm, and the expertise of the operator in assigning specific clusters to definite crop classes.

In the 1996 project, we have tested the performance of a classification approach based on the use of synthetic channel composites (Nezry et al, 1996). In a synthetic channel composite statistical parameters of the time series are combined in a 3 channel composite. In our case, we have used the mean of the series, the range (difference between maximum and minimum) and the date of maximum backscattering. Mean, range and date of maximum are calculated on a per pixel basis. Classification was then based on an isodata clustering followed by regrouping of clusters and class assignment (GAF et al, 1997).

The evaluation of the classification performance with information from dedicated field surveys (SYNOPTICS, 1996b and 1996c) and MARS Action 6 surveys revealed that the results were non-optimal (SYNOPTICS, 1996d). This was partly due to several factors that influenced the quality of the individual images (geocoding, speckle filtering, adverse surface conditions), but also due to difficulties in the interpretation of the synthetic channel composite information content. A main drawback of the applied synthetic channel composite technique is the effect of soil moisture on the range and date of maximum backscattering. In the classification, one is mainly interested in separation of roughness and roughness transition events. In fact, soil moisture induced backscattering change constitutes an unwelcome noise factor that needs to be compensated first.

The evaluation led to an improvement of a method proposed during the study phase in 1995 (SYNOPTICS, 1996a). The so-called byte-sliced composite method recognises the fact that in typical SAR imagery for the autumn and spring periods, delineation of classes of agricultural fields is typically limited. The method codifies this observation by interactively slicing the SAR image (calibrated filtered geocoded (CFG) data) into 4 classes. In this way, an image can be stored with 2 bit per pixel. A logical OR operation on bit-shifted images creates a composite in which the indices are directly related to the membership of the various slices in the individual images. Since the limits of the slices are known, a direct correspondence with backscattering coefficient limits exists. The procedure to generate byte-sliced composites is illustrated in Figure A, for an combination of 3 images. The result is a composite in which the indices range from 0 to 63 (4³-1). Extension into integer or even long-integer composites is possible.

The use of the composite has the following advantages:

• Up to 4 images can be combined in one byte (up to 8 in an (unsigned) integer, etc.)
• The effect of soil moisture is compensated partly by the introduction of the histogram quantile stretch, assuming a multiplicative effect of soil moisture change on the backscattering coefficients of bare soil
• Since all processing steps are documented, a full retracing of the procedure is possible.
• The calibrated signature values can be used in a signature data base. Automation of the interpretation procedure may be realised.

Figure A. Byte-sliced composite generation for SAR image interpretation.

The image composite is input into an image interpretation scheme, that is not very different from other image classification exercises (Figure B). Apart from the composite a number of other ancillary image products are used e.g. previous classification results, optical imagery, digital elevation models and raster maps as well as non-image data (meteorological records, crop calendar information (Kidd et al, 1997)). All these products are combined in support of interpretation of the temporal signatures that lead to the generation of a certain index in the byte-sliced composite. For instance, if one has a series of 3 images from autumn and spring, then index 0 in the composite means that the pixel (or field) has been assigned to the lowest slice in all 3 images (and is grassland in this case). Index 63 means membership of the three highest slices for all instances.

For autumn and spring combination, membership to slices is mainly determined by surface roughness conditions, which determine the backscattering coefficient of the bare and partly vegetated fields. For combinations that include summer imagery as well the characteristic signature of the crop canopy is another determinant.

Figure B. The workflow in the image interpretation procedure.

RESULTS.

The use of the byte-sliced composite is illustrated with an example ERS time series of the Great Driffield MARS site in the United Kingdom. This site is characterised by a large area of arable land in the centre of the site, with winter and spring cereals, oil rape seed, potato and sugar beet as main crops. In the lower areas of the site, grass land is found more frequently. The site is located on the East coast, with its centre at approximately 54 N and 0 W. Annual precipitation amounts to some 600 mm and temperature extremes are typically between -5 and 25 degrees.

In Figure C, the results for a combination of 3 winter and spring images is shown (26 December 1995, 25 February and 31 March 1996). In this image, grass land, winter crops and spring and summer crops are delineated. The differentiation between spring and summer crops is the most difficult, since the spring of 1996 was unusually cold, with snow covers still present in half March. The classification results were assessed with field survey data that were collected in the second half of March (SYNOPTICS, 1996b). An accuracy of approximately 70 percent is reached, with apparent confusion between grass land and winter crops. This confusion is due to the spectral proximity of these two

Figure C. Crop classification result for the Great Driffield MARS site, using autumn and spring ERS SAR images.

classes (which are effectively smooth bare fields, with little vegetation). In Kidd et al, 1997, we have presented the classification of a 4 image composite (25 February, 23 April, 9 June and 14 July 1996). In it, at least 6 distinct classes can be separated. The distinct signatures and contrast in the summer images cause a crisp delineation of field boundaries. The most interesting features of this combination are the distinction between winter wheat and winter barley, and the classification of oil rape seed (these are the major crops in the area that are supported with EU subsidies). In fact, the choice of the image combination was tuned to this purpose. Again, the classification was quality assessed with the ground survey data that were collected in March and those of dedicated surveys on cereals in June and July (SYNOPTICS, 1996c). The overall accuracy is around 80 percent (for details, see Kidd et al, 1997). The winter cereal classification has an accuracy higher than 90 percent. The separation between winter barley and winter wheat is possible since cereal ripening has a distinct effect on the crop backscattering signature, and barley ripens approximately 3 to 4 weeks earlier than wheat. The June and July images are exactly timed around this period. The separation of oil rape seed is possible, because the backscattering strongly increases after the end of April. Most summer crops (potato, sugar beet) also have a relative high signature in summer, but development is somewhat later, and the February image is used to separate these from oil seed rape because of their rough (ploughed) surface conditions. The inclusion of this early spring image is also essential in the separation of grassland.

A comparison of the classification results with SPOT derived results could only be made with the intermediate classification results for April, May and June imagery. This has shown that the SAR derived product performs better, especially with respect to the delineation of distinct crop classes. In the SPOT imagery, crop class specification is often limited to generic classes, even after the (third) acquistion in June. The separation between winter barley and winter wheat seems not to be possible in SPOT imagery. On the other hand, early spring SPOT imagery is very useful in fine-tuning separation between bare soil, winter crops and grass land. Limited synergetic efforts show positive effects of combining this image with early season SAR series. Also, field delineation is usually much better in single data SPOT imagery. Combinations with later SPOT images do not show significant improvements, unless the SPOT image captures a distinct spectral event, such as the flowering of oil seed rape at the Bernburg site in early June.

A further example is shown in figure d. The classification result for the Bernburg MARS site in Germany is shown. The composite method was extended to integers to create compositions of 6 sliced-images. Note that this creates 4096 different indices that need to be combined in clusters. The set of indices is dominated, however, by a subset of indices that relates to the signatures of distinct crop classes. In fact, some 200 indices constitute more than 65% of the pixels, with another 23% pixels included as masks. Thus, the majority of the indices belong to small fragments, that somehow need to be clustered. In the example image, these small fragments have been left unassigned to illustrate this point.

SAR image combinations in general are very effective in highlighting effects that are of use in image interpretation exercises. For instance, the separation of spring cereals can be improved with additional images in early April and May. An example combination that shows the effect of the soil is shown in Figure E. The image is a composition from the March 31 and April 9 acquisitions. During this period, there was no rain. With relatively little vegetation development because of low temperatures, the composite shows the effect of differential drying in the scene. A comparison with the 1:1,000,000 soil map of Europe is shown. The effect of drying is largest (4-6 dB) for the loamy soil in the centre area of the site, and much smaller for the lower areas that are characterised by clayey soils which suffer from water logging. Note also that roughness changes are included in this image, but the clear drop in backscattering values for winter cereals indicate that soil moisture change is the main mechanism (vegetation coverage is below 25% at the time). Other combinations are useful in emphasising roughness effects. For example, an ascending and descending

combination in April clearly show the effect of look modulation for rough field with clear row patterns (ploughed fields and potato ridges).

Figure D. A byte-composite classification for the Bernburg site (6 image combination).

Figure E. A SAR image combination (left) for the Great Driffield site spanning a short period of drying in early April. The image is compared to the soil map in the right image.

Another aspect of signature analysis is the use of reference fields that are used to compensate for soil moisture change. Winter crop fields in the autumn and early spring, and potato fields up to the end of June serve this purpose very well. We have used this approach to highlight backscattering features that are related to roughness change and vegetation development. This shows that the major trends for the Great Driffield and Bernburg sites are very similar, though different in absolute levels. We expect this to be related to both soil and crop canopy differences. In an experiment for winter cereals, we have used the different sensitivity to soil moisture change in summer for the Great Driffield and Bernburg sites to highlight canopy attenuation effects that clearly differ between the two sites (SYNOPTICS, 1996f). It is speculated that this effect is due to lower wet biomass of the cereal canopy at the Bernburg site because of soil water deficit.

CONCLUSIONS.

We have presented a simple technique for deriving crop classification products from multi-temporal SAR series. The method is based on the fact that in single SAR imagery only a limited amount of backscattering classes can be delineated, especially in the autumn to spring period. A recombination of scattering classes leads to a better separation of image clusters that are directly related to distinct backscattering signatures. The interpretation of these signatures into crop classes is then driven by the use of ancillary data resources.

We have shown that the performance of the proposed method is satisfactorily for separation of generic crop classes in early season imagery. Performance with image combination that include summer acquisitions was shown to be better than those currently produced in the optical image interpretation exercise in the MARS Action 4 project (at least for the Great Driffield site). Especially the more reliable timing of SAR image interpretation exercises is of considerable importance in the context of crop area estimation. Furthermore, the method can easily be extended to include any combination between 2 and 16 images (although a selection of 6 seems to set a practical limit).

The rigorous documentation requirements of the method allow a full reproducibility and evaluation of processing steps. Since all intermediate products are stored, small errors are easily corrected.

Other image recombinations were shown to be helpful in delineating backscattering features that are useful for determination of soil and vegetation cover effects. Such features can then be used in the assessment of biophysical parameters or specific crop types.

It should be emphasised that the application of the method requires precise geocoding of SAR imagery and a good performance of speckle filtering and image calibration. Also, an image pre-screening exercise is necessary to filter out images with sub-optimal class delineation characteristics (e.g. due to snow, frost or drought). Obviously, successful implementation of a agricultural monitoring exercise with SAR data requires close interaction between the data supplier and the data user. Also, pricing policy should be tailored to wide area coverage and multi-temporal series in order to generate statistically significant crop sampling. Decentralised image reception on location could very well enhance the acceptance of SAR imagery in such agricultural monitoring exercises.

Since the MARS project constitutes an advanced operational framework for remote sensing applications in agricultural monitoring, ample possibilities exist to carry out further research in SAR dat use. In fact, the various SAR activities has led to the creation of a formidable data set. Future efforts will be directed to integration with optical data sets, improvement and automation of the interpretation cycle and testing of wide area sampling.

ACKNOWLEDGEMENTS

This work has been partly carried out under contract OJ 95/C203/07 which was funded by DG VI and supervised by the MARS project.

REFERENCES

GAF, SYNOPTICS and NRSC, 1997, A pilot project on the use of active microwave sensors for the rapid area estimation of Agricultural crops, Final Report O.J.95/C203/07, 700+ pp.

Kidd, R., Lemoine, G., and de Groof, H., 1997, Integration of ERS SAR Classification Products in the MARS Activity B "Rapid Area Estimation" Methodology, Proc. 3rd ERS Scientific Symposium, Florence, Italy, 17-20 March 1997.

Lemoine, G., van Leeuwen, H., and de Groof, H., 1995, Proc. ERS Applications Symposium, London, Dec 4-8, 1995.

Nezry, E., Solaas, G., Remondiere, S., and G. Genovese, 1995, Mapping of next season's crops during the winter using ERS SAR. Earth Observation Quarterly, No 50 Dec 1995.

NRSC, 1996, Evaluation of the SAR Processing Chain. Project Report O.J.95/C203/07, National Remote Sensing Centre, Ref DG-TR-NRL-AP-001, Issue 1.0, July 1996

Sowter, A., Lemoine, G., and de Groof, H., 1997, , Proc. 3rd ERS Scientific Symposium, Florence, Italy, 17-20 March 1997.

SYNOPTICS, 1996a, ERS-1 Data Analysis, Final Report MARS Contract Study 10946-ISP NL.

SYNOPTICS, 1996b, Spring field survey report, Project Report O.J.95/C203/07, Ref O.J.95/C203/07/1.1 /GGL.

SYNOPTICS, 1996c, Cereal field survey report, Project Report O.J.95/C203/07, Ref O.J.95/C203/07/3.1 /GGL.

SYNOPTICS, 1996d, Methodological Advancements, Project Report O.J.95/C203/07, Ref O.J.95/C203/07/5.1 /GGL.

SYNOPTICS, 1996e, Assessment of Crop Separability, Project Report O.J.95/C203/07, Ref O.J.95/C203/07/5.1 /GGL.

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