INTEGRATION OF ERS SAR CLASSIFICATION PRODUCTS IN THE MARS ACTIVITY B &QUOTRAPID AREA ESTIMATION&QUOT METHODOLOGY

ABSTRACT

The MARS Activity B "Rapid Area Estimation" has been an operational project for the last 5 years. It is based on the interpretation of high resolution optical imagery, collected throughout the growing season, for 60 sites in the EU to derive estimates on crop area changes at the level of the European Union. Within the MARS Agricultural Information System, a number of ancillary data products are routinely collected which may aid the image interpretation. This paper addresses the use of these ancillary products in support of the interpretation of ERS SAR time series for area estimation. The work has been carried out as part of a pilot project on the use of ERS SAR data for rapid area estimates of agricultural crops which was initiated by the DG VI -Agriculture of the European Commission under the scientific and technical supervision of the MARS AIS unit of the Joint Research Centre. While the analysis of backscattering signatures has been presented in a separate paper, we concentrate on the methods used for, and constraints placed upon the use of ERS images for crop area estimation. Methods for the fusion of ERS data with other spatial and non-spatial data sets in a dedicated information system are presented. Limitations and constraints of the different data sets are documented and recommendations made. Although currently carried out in a separate data analysis environment, we show that integration in a dedicated interpretation software environment is highly feasible.

Keywords: Agriculture, crop classification, operational, MARS project

INTRODUCTION

This paper offers an outline of the MARS data sets as used in the context of the DG VI project "A Pilot project on the use of active microwave satellite remote sensing data for rapid area estimation of agricultural crops". After a brief overview of the MARS project, the application and incorporation of these data sets into a SAR classification methodology is then presented, along with the inherent limitations, before finally drawing conclusions from the use of, and recommendations for the future implementation the MARS data sets.

The MARS Project (Monitoring Agriculture by Remote Sensing ) has been operational since 1988 when it was launched as a pilot project for remote sensing applied to agricultural statistics. The overall operational aims of MARS project,(MARS-STAT) (Vossen, 1994) are twofold: firstly to distinguish, identify and measure, on a yearly basis, the area of economically important crops in Europe. Secondly, derivation of estimates of crop production early in the year. This rapid estimation of agricultural crop area is implemented within the framework of Activity B, (Action 4) of the MARS project. These estimates are presently derived from analysis of high-resolution optical satellite images acquired either from SPOT or from Landsat Thematic Mapper.

Action 4 is based upon the derivation of crop area estimations at a European level, with results extrapolated from a series of 60 sites. Each of these sites, distributed over the member states of Europe, covers an area of 40 x 40 km, thus offering a total coverage of 4% of the agricultural area within Europe.

The accurate identification of crop types depends on the availability of images acquired within specific time windows throughout the crop growing season. The timeliness and the quality of information derived from passive optical systems suffers from a number of constraints, for example extensive cloud cover in Northern Europe throughout the agricultural season. It is also noted that the discrimination of crops early in the growing season based solely upon their radiometric signatures in the optical domain, is inherently limited.

The on-board Synthetic Aperture Radar (SAR) systems of ERS satellites are, to all extents, independent of cloud cover and daylight conditions. These capabilities are of particular interest for the MARS Action 4 test sites lying in Northern Europe, where the ability to collect optical imagery within relatively narrow time windows is problematic.

But along with this SAR has a great potential for early crop identification. The sensitivity of this instrument to changes in surface structure and moisture, allow for the identification and classification of soil surfaces being prepared for different crop types during autumn and winter. Therefore allowing a possible approach to an early identification of crops and, consequently, crop areas.

During the course of its activities the MARS Project has collected, and archived large amounts of data and gained experience in the fields of remote sensing, agronomy and related statistical analysis. These data sets held within the MARS Project vary from remotely sensed data, agronomic data, meteorological data along with derived products, and other types of spatial and statistical material.

In the context of the DG VI project "A Pilot project on the use of active microwave satellite remote sensing data for rapid area estimation of agricultural crops" several of the above mentioned data sets of the MARS project have been utilised in the development of the SAR interpretation methodology. Although this project was only aimed at providing area estimations for 20 Action 4 sites within Northern Europe, the rationale behind this paper is that the methodology for the use of these data sets can be applied to all Action 4 sites.

DESCRIPTION OF DATA SETS

In this classification methodology six unique sources of ancillary data have been identified. Specifically these are high resolution optical data products, meteorological data, digital elevation models, soils information, agronomic data sets, and ancillary site information.

High resolution Optical Products

For all of the 60 Action 4 sites of MARS Project, high resolution optical imagery (SPOT and Landsat TM) are acquired up to 4 times a year. The analysis of this imagery forms the foundation of the MARS Project activities in the area of yield prediction. The output from this analysis results in two data products, per image, at a site level. Firstly the unclassified optical imagery, and secondly a visually interpreted classification of the optical imagery. Over the operational lifetime of the MARS project the MARS archives now holds approximately 1750 raw optical images, along with all corrected and classified imagery.

The optical imagery undergoes a complete correction procedure. For each image radiometric, atmospheric and geometric correction is applied. In order to minimise data handling procedures data extraction is required. For SPOT XS images all 3 channels are extracted, whilst for TM images only channels 2,3,4 and 5 are extracted. The final product is then resampled to 20m pixels. (Noting that the final product is orientated to the SPOT imagery.)

The classified, interpreted, imagery is available at two levels. The main product is a final composite classification, which has been updated and corrected with ground survey observations, and is available on a yearly basis. In the classification procedure a series of processing steps are applied including the application of specific masks for non-agricultural areas. Secondly the intermediate visual interpreted classification is available for all optical images.

Meteorological Data

The MARS Project meteorological database contains historical daily weather observations from several hundred meteorological stations across Europe. The spatial extent of this data includes coverage of the new EU member states. This data set is updated on a decadal basis, with a hiatus of 4 days for processing and validation of data This is then extrapolated to give meteorological data scaled to a 50 by 50 km grid coverage of Europe, as implemented in the CGMS (Crop Growth Monitoring System) (van der Voet et al, 1994) system of the MARS Project. This data set is available either as diurnal measurements or decadal (10 day) averages. In order to exploit this information each site of interest must be related to the 50 by 50 km grid coverage of Europe.

Digital Elevation Models

Digital elevation models (DEM) exist in two formats. The first is a set of low resolution DEMS, as used in the within the GRIPS (Geographic and Radiometric Image Processing System) software for the pre-processing of the optical high resolution imagery. These offer the complete coverage of each Action 4 site, and have pixel resolutions in the order of 160 to 640 m.

Further to these DEMS a series of high resolution (50 m resolution) DEMS exist clipped to each Action 4 site.

Soils Information

The MARS Project activities include the development and maintenance of the European Soils Database (via the European Soils Bureau). This database is presently available for 12 Member States of the EU, and also covers extensions into Central and Eastern Europe. For the remaining three EU member States work is being scheduled

The European Soil Database (King et al, 1995) is based on the 1:1,000,000 FAO soil map, and consists of soil mapping units (SMU) as its basic polygons. The SMU's describe the basic soil characteristics in terms of physiographical parameters (e.g. geological parent material, altitude, land use and limiting factors). Every SMU can have several soil topological units (STU), which correspond to soil types. Soil types are given using the FAO soil name. For a number of STUs detailed information on soil texture, slope and agricultural restrictions are available.

Agronomic Data Sets

The MARS Project Action 6 yield and ground survey data is a point-based system that has been operational since 1994. The sampling points are located in a grid of segments that are distributed across the sites. The positioning of the segments depending upon agricultural stratification of the site. These segments are 1440m by 1440m and contain 40 points distributed on a grid. For each point, the crop class, yield, and cropping information is registered on a yearly basis during a summer survey. Each site has a maximum of 16 segments, giving a maximum total of 640 points of data per site.

This point based agronomic data set is also reinforced by documentary agronomic information as compiled for use within CGMS.

Ancillary Site Information.

This collated data set consists of geographic information at a site level, detailing site and point locations, locations of sites with respect to European grid coverage, and image projection information.

MANAGEMENT OF DATA SETS

In the scope of this project there is obviously a necessity to provide effective management and access to all the ancillary data sets and data products. To realise this requirement a structured coherent data base has been developed and implemented (SYNOPTICS, 1996a). Along with the ancillary data sets from the MARS archives, a number of other data sets, collected within the context of this project are also stored. Specifically these include the results from the dedicated field surveys carried out by SYNOPTICS in the spring and summer of 1996 (SYNOPTICS, 1996b and SYNOPTICS, 1996c), and extracted (polygon based) signatures from ERS and SPOT imagery (SYNOPTICS, 1996d and SYNOPTICS, 1996e)

Within this database data for 20 sites is included. Field survey data for 1996 comprises 5000 Action 6 point locations, nearly 1900 polygons for 8 sites surveyed in the spring and over 400 polygons for the Bernburg and Great Driffield sites in summer. Over 25,000 ERS backscattering coefficients have been included for the various ground locations.

The 47 data base tables have been organised into 7 thematically related groups, as listed in Table 1. By structuring all relevant site information in this way, access and retrieval of information has been vastly improved.

Table 1. Organisation of thematic groups in data base.

INCORPORATION OF DATA SETS IN THE ERS METHODOLOGY

All of the above mentioned data sets have particular roles in the processing and classification of multi-temporal SAR imagery. In order to emphasis the specific uses of the data sets, the two aspects of the processing (Figure 1) and classification methodology (Figure 2) are approached and presented separately. The specific uses of these data sets are summarised in Table 2.

Data Processing

Data screening is performed to ensure the quality of any imagery before the product is ordered. Initially temporal windows of opportunity are identified for dates of optimal image acquisition, (optimal in the sense of the possibility for identification of distinct crop types). Imagery is then selected on a basis of complete site coverage occurring during the optimal dates, and a quick look of the image acquired. Meteorological data for dates preceding image acquisition is required to check for any meteorological events (heavy rain occurring at time of acquisition, frost, snow, etc.) that are likely to influence the quality of the imagery.

As the final composite (SAR) image is only required for the area extents of each specific site all images are subset to the Action 4 site and projected in SPOT image coordinates.. This requires the use of precise site location and rotation knowledge. Within this project all image data has been processed with the dedicated TSAR (Topographic SAR Processor) system at NRSC Ltd. (NRSC, 1996) using the 50 meter spaced DEM for detailed calibration and terrain correction. The data used in this study are derived from calibrated, geo-referenced GMAP (Nezry et al, 1993) filtered PRI data (termed CFG data in this report).

Figure 1. The data processing procedure

In the production of the Composite Imagery , (byte sliced image composite, Lemoine et al, 1997), it is necessary to have an understanding of the site specific agricultural conditions at the time of acquisition. Firstly to remove areas of potential confusion in the image non agricultural areas can be masked out using the masks derived from the optical classified imagery. Information with respect to the crop types and practices in the area can be determined from both agronomic data sets and from the explicit classes identified in the final verified optical classifications. By incorporating this knowledge with the historical agronomic information, crop rotation schemes can be determined, and hence the likelihood of finding certain crop types in certain areas can be evaluated. The meteorological data is here used not only to determine the likelihood of extreme or adverse conditions, but also, in conjunction with the soils data sets, for estimation of the surface moisture conditions at the time of image acquisition.

Classification Methodology

The classification procedure applied is analogous to the well known unsupervised classification methodology. It is imperative that within this procedure each interpreter working on a per site basis has as much relevant information as possible. By using these data sets the interpreter must develop a complete understanding of not only the topography of the area, but also how this relates to the expected agricultural practices at the time of image acquisition. Furthermore the interpreter must have a complete understanding of the physical nature of the cover types present, and also an understanding of the effects of agricultural practices upon the backscattered signal.

Figure 2. The workflow in the Classification Procedure

Table 2. Summary of use of data Products

In areas where there is a lack of a priori knowledge about certain cover types, it is also possible to incorporate outputs from agronomic models, such as the WOFOST crop growth model (Supit et al, 1994), to predict cover development stages.

A summary outlining the main uses for the MARS derived data sets is given in Table 2.

RESULTS.

To illustrate the benefits of the inclusion of these data sets in our classification methodology, we present in Figure 3 a comparison between our knowledge steered classification and an unsupervised isodata classification. Data is shown for a 20 by 20 km subset of the Great Driffield MARS Action 4 site in the UK. Four CGF images were used (dates: 25/02/96, 23/04/96, 09/06/96, 14/07/96). The timing of this series of images is considered optimal for the discrimination of the main crop types across this site (a total of 13 images between November 1996 and August 1996 were available).

In our classification the following five distinct classes can be delineated: oilseed rape (RAP), spring barley (SBA), grass land (GRA), winter barley (WBA) and winter wheat (WWH). For the summer crops, a distinction can be made between the peas/sugar beet (PEA/SBT) fields and other summer crops (SUM, mainly potato (POT), spring oil seed rape).

The classification accuracies have been assessed with the field survey results for spring and summer. The spring survey, carried out at the end of March concentrated on separation of winter crops and early spring and summer crops (the latter two both at bare soil stage). The summer survey, carried out in mid June and Mid July was focused on separation of cereal types. The results are summarised in Table 3 and Table 4.

Table 3. Classification accuracy assessment with spring field survey data. The top table is for the ISODATA classification, the bottom table for the byte-composite method.

In the tables, the polygons that are rightly classified are printed boldface, those that are wrongly classified are underlined. Ambiguous cases are in normal typeface. The total number of fields are 187 and 186 for the ISODATA and byte-composite classification, respectively. The total number of polygons was 205 (a 30% threshold was used, and small classes have been removed). The number of rightly classified fields for the ISODATA classification is 124 (66.3%), the number of wrongly classified fields 53 (28.3%). For the byte-composite these number are 130 (69.9%) and 47 (25.3%) respectively.

Looking at confusion between spring barley and winter wheat is the most obvious. This is most likely due to the late timing of the available April acquisition. It seems more appropriate to aggregate the classes SBA and WWH. In that case, the number of rightly classified fields increases to 139 (74.3%) and 148 (79.6%) for ISODATA and byte-composite classification respectively.

Table 4. Classification accuracy assessment with summer survey data.

In Table 4 the results for the summer survey are shown. A total of 208 fields were sampled. 199 field were used in the assessment for both methods. The number of rightly classified fields is 129 (64.8%) and 145 (72.9%) for ISODATA and byte-composite methods respectively. Aggregation of SBA and WWH yields 156 (78.4%) and 172 (86.4%).

Although the overall performance improvement of 5% seems not too impressive, other observations point out a generally better performance of the byte-composite classification method. For instance, the various summer crop types are better delineated in the byte-composite classification. This is due to the fact that cluster assignment is better controlled in the byte-composite method. During signature analysis and interpretation, special weights can be given to backscattering features that relate to a certain crop type, while ISODATA classification lacks this option. Of importance also is that we have used a priori information to tune the number of clusters in the ISODATA classification. Thus, also simple ISODATA classification benefit from the use of ancillary data.

As a last quality assessment step, we have generated area estimates for the complete site for both classification, and compared these to the results generated with SPOT data (3 images) during the Action 4 activity in 1995 (Table 5). There are a number of interesting observations in this table. The first is related to the definition of the classes in the SPOT and SAR-based classifications. The SPOT classification includes fallow land and fodder. Since separation of this class is not possible in the SAR classification, we expect this class to be aggregated in the GRA class for SAR (considering that surface conditions are approximately equal). The ISODATA classification might have grouped these in the class OTH. The difference in WWH percentages is found back in the SBA class for the SAR. Apparently, neither SPOT nor SAR is successful in separating these classes. RAP separation is by far superior in SAR data than in the SPOT data. POT stands out better in SPOT data. The SPOT classification includes 13.3% of mixed classes, while the byte-composites class has only 1.8% unknown pixels (which are mainly distributed randomly across the site).

Table 5. Area estimates derived from various classification products for the Great Driffield site.

CONCLUSIONS.

We have presented the use of a SAR classification methodology within the operational framework of the MARS program. The methodology is based on a sound analysis of SAR backscattering signatures with the support of a large set of ancillary data sets that are part of the MARS archives. The relevance of the different data sets has been described in detail.

In the SAR processing methodology, ancillary data sets already play an important role at data planning (crop calendars, previous classification results) and data screening (meteorological data, soil maps, crop maps). During interpretation this set, together with site knowledge, significantly improves the accuracy of cluster assignment.

The method was illustrated in comparison to an unsupervised ISODATA classification (which was partly guided by a priori information). An improvement in classification accuracy of approximately 5 percent was found for the example SAR series for the Great Driffield site. A comparison of area estimates derived from both byte-composite and ISODATA classifications against SPOT 1995 results, suggest that the byte-composite method can generate crop area statistics that are better than those derived from SPOT imagery. It should be noticed that although the results presented in this paper are for one site only, consistent improvement have also been found for the Bernburg site (Germany).

Confirmation of these results would need extrapolation to other MARS sites. This is currently done in a follow-up project, funded by DG VI.

With respect to improvement in ancillary data sets, we have found that ground survey data on a field polygon basis are more advantageous in classification accuracy assessment than Action 6 ground survey data (point based) that are part of the MARS archive. Soil maps sometimes show a shift in location, due to the rather large scale. At the moment, the MARS project team is considering the creation of a 1:250,000 soil map data base. Crop information might be further improved with information on tillage methods and their effect on surface conditions.

Ancillary image sets might include segmented optical images for better field boundary detection and topographical maps for better registration of SAR imagery at geocoding stage.

Figure 3. Colour coded classification result for Great Driffield, using an optimal 4date composite. Knowledge steered classification (left image), and an unsupervised ISODATA classification (right image).

ACKNOWLEDGEMENTS

This work has been carried out under contract OJ 95/C203/07 which was funded by DG VI and supervised by the MARS project. The authors would like to thank Iwan Supit and Carmelo Attardo of the MARS project for critical contributions to this work. Further appreciation goes out to SYNOPTICS staff Leon Schouten and Onno Luimstra for carrying out the field surveys and Hans van Leeuwen for comments and support.

REFERENCES

King, D., Burrill, A., Daroussin, J., Le Bas, C., Tavernier, R., van Ranst, E., 1995: The EU Soil Geographic Database, European Land Information Systems For Agro-Environmental Monitoring, Eur 16232 EN, pp 43-60

Lemoine, G., Kidd, R., van Leeuwen, H., de Groof, H., 1997, Methodological advancements in using ERS SAR data for crop area estimation,. In Proceedings 3rd ERS Scientific Symposium, Florence, Italy, 17-20 March 1997

Nezry, E, 1993, The Refined Gamma-Gamma MAP Speckle Filter, EEC Expert Contract n°EARS 92-0004-FR, pp74, 1 June 1993

NRSC, 1996, Evaluation of the SAR Processing Chain. Project Report "A pilot project on the use of active microwave sensors for the rapid area estimation of Agricultural crops" O.J.95/C203/07, National Remote Sensing Centre, Ref DG-TR-NRL-AP-001, Issue 1.0, July 1996

Supit, I.., Hooijer, A.., A.., van Diepen, C., A., 1994, System Description of the WOFOST 6.0 Crop Simulation Model Implemented in CGMS, Volume 1: Theory and Algorithms. An Agricultural Information System for the European Community, EUR 15956 EN, pp144

SYNOPTICS, 1996a, Site Data Base Description, Report O.J.95/C203/07, Ref O.J.95/C203/07/7.1 /GGL.

SYNOPTICS, 1996b, Spring field survey report, Project Report "A pilot project on the use of active microwave sensors for the rapid area estimation of Agricultural crops" O.J.95/C203/07, Ref O.J.95/C203/07/1.1 /GGL.

SYNOPTICS, 1996c, Cereal field survey report, Project Report "A pilot project on the use of active microwave sensors for the rapid area estimation of Agricultural crops" O.J.95/C203/07, Ref O.J.95/C203/07/3.1 /GGL.

SYNOPTICS, 1996d, Assessment of Crop Separability, Project Report "A pilot project on the use of active microwave sensors for the rapid area estimation of Agricultural crops" O.J.95/C203/07, Ref O.J.95/C203/07/5.1 /GGL.

SYNOPTICS, 1996e, Potential of Data Combinations, Report O.J.95/C203/07/6.1 /GGL.

van der Voet, P., van Diepen, C.