CLASSIFICATION OF ERS-1 SAR DATA OVER SEVILLE (SPAIN) FOR AGRICULTURAL STATISTICS

ABSTRACT

A time series of nine ERS-1 SAR images of the MARS sample site Seville, four SPOT XS images and ground surveys of 17 MARS segments all from 1992 have been available for the study. Preliminary masks of non-agricultural areas derived from thematic maps are refined by multi-temporal, unsupervised classification of optical satellite data. Stratification was performed using visual interpretation of SPOT data for refining macro-structural units of the region around Seville extracted from relevant literature. During preliminary investigations it turned out, that significance of supervised classification is limited due to limitations of the available ground truth. A robust, transferable approach is presented, which allows to retrieve crop types/single crops from ERS-1 SAR data on the base of unsupervised clustering. The recombination of clusters to classes is performed using ancillary knowledge (e.g. local agricultural practice, parcel structure). The method is detailed with the example of a single stratum. Application to the other strata of the Seville site proved functionality of the method. There are significant indications, that ERS-1 SAR data could be used as complement or substitute to optical satellite data. With regard to the operational needs of the MARS project the potential for an operationalisation of the approach seems to be high.

Keywords: ERS SAR, MARS, Masking, Stratification, Unsupervised Classification

1. INTRODUCTION

The MARS (Monitoring Agriculture with Remote Sensing) project includes an activity to derive area statistics for agricultural crops based on satellite imagery. The Joint Research Centre (JRC) decided to select more than 50 sites spread over Europe as an area frame sampling scheme. For the sample sites each of 40 km 40 km size SPOT or Landsat TM data are acquired nominally four times during a growing season (MEYER-ROUX, J. and VOSSEN, 1994). The main drawback consists in the fact, that cloud cover often hinders acquisition of optical images. ERS Synthetic Aperture Radar (SAR) data will allow to fill temporal gaps of optical coverage, as well as to ensure all over the year coverage of the sample sites. Considering potential of ERS SAR data for agricultural use (e.g. KOHL et al., 1994; HARTL et al., 1995), it is intended to introduce these data into the MARS project in order to

• get complementary or additional information to optical data

• substitute optical data, especially in the early growing season

in the framework of MARS rapid estimates (RE). For a corresponding study the area around Seville, one of the important agricultural regions of Spain, with a wide range of crops, was selected. The scientific and practical results, achieved with this study, should be made transferable to other sample sites and to operational applications like RE aiming at a possible integration of ERS-1 SAR data into the operational system of MARS.

2. TEST SITE

Most of the Seville testsite belongs to the lowlands of Andalusia, which are characterised by the Guadalquivir valley. Morphologically this landscape is slightly hilly (Campinas) except of some steep slopes close to the Guadalquivir. Between the southern part of Seville and the delta of Guadalquivir flat marshland (Marismas) is located. Corresponding to MADUE&NTILDEO (1991) the region around Seville can be stratified into 13 different macro-structural agro-physical units. Only 11 out of 13 units are covered by the area under investigation.

In principal climate in Andalusia is typical Mediterranean with hot, dry, summers and mild winters. Due to few rainfall during summer months this area belongs to the semiarid region. This implies, that either irrigation or dry farming is practised for successful agriculture. The mosaic of soils in the area of interest is rather different. Main soil types are Tirs soil, recent Terra Rossa, relicts of Terra Rossa with lime and marl, Marsh soils.

Grown permanent cultures in Andalusia are to olives, wine, fruits (citrus, peaches), eucalyptus, pine trees, pastures, horticulture. Grown temporary crop types in Andalusia wheat, barley, oats, sunflower, maize, sugar beet, safflower, soya, sago, sorghum, potatoes, rice, cotton, broadbeans, chickpeas, water- and sugarmelons, tomatoes, lucerne, herbs (MADUE&NTILDEO (1991).

There are two cultivation practices of temporary crops:

• Traditional dry-farming,
• so called wet farming (artificial irrigation).

Dry farming (two or three field rotation with a long period of fallow) has decreased within the last years. Traditional crops for dry-farming were cereals, maize, and crops grown on fallow land like broadbeans, chickpeas, and other fodder plants. If wet farming is practised, rice, sugar beet, potatoes, and legumes are grown but with minor importance. Because of EU-activities some crops were added to the traditional crop catalogue like sunflowers, soya, and sago. Therefore the traditional crop rotation has also changed.

3. DATASET AN PRE-PROCESSING

Satellite and ground truth data delivered by SAI/AIS of JRC (Space Applications Institute/Agricultural Information System) have been implemented on the computer system at the Institute for Navigation (INS) of Stuttgart University. Additionally topographic and thematic maps of the Seville area as well as statistical and geographical information have been acquired. For the Seville MARS sample site there have been chosen 17 instead of regularly 16 MARS segments. This 17 segments were visited by field survey teams during May and June 1992. Tab. 1 contains a list of the available ERS-1 SAR PRI and SPOT XS acquisitions.

Tab. 1: Acquisition date of Satellite data

Additionally a SPOT-PAN scene of 28.02.92 was added giving an overview and orientation in fine resolution, and allowing photo-interpretation of details (Fig. 1).

Pre-processing for ERS-1 SAR PRI images partly applied at SAI/AIS partly at INS comprises extraction of subframes, image resampling to 20 m 20 m pixel size, compression of 16- to 8-bit data, speckle filtering (LOPES et al. 1993), and full calibration (LAUR, 1992).

The available SPOT-PAN ortho-image was used as master image for the geocoding of the ERS-1 SAR images. Due to the use of an ortho-image, i.e. a map representation at sea level as master image it is not absolutely necessary to introduce a DEM into the geocoding process. During geocoding by polynomials of n-th order and during following resampling, orbit and relief introduced distortions are corrected. A set of 45 ground control points and a first order polynomial was used for this procedure resulting in standard errors of about 100m in East and about 50m in North direction.

4. POTENTIAL OF ERS-1 SAR DATA FOR CLASSIFICATION

A test has been performed based on means and standard deviations of training areas (DOBSON et al., 1992) in order to assess the potential of supervised classification using the polygons of MARS segments as training areas. Results show, that only very limited differentiation of landcover and crop types is possible. According to a detailed examination also most of the histograms showed up multimodal distributions, i.e. several spectral classes within one ground truth class. It also turned out, that the number of samples available for a crop or landcover type was not sufficient in many cases with respect to statistical (i.e. minimum number of samples >100) or physical considerations.

Due to these restrictions it was necessary to modify the approach in the sense, that first a statistically sound procedure has to be applied to identify the intrinsic information content ERS-1 SAR data, followed by an appropriate input of ancillary information for a final classification (NEZRY ET AL., 1995). The following main prerequisites for this approach have been identified and considered to be fulfilled for the Seville dataset:

• Homogeneous scene conditions (as provided by masking and stratification see chapters 5 and 6),
• knowledge about occurring land use / crop classes,
• knowledge about phenological development and related spectral reflectance / backscatter behaviour of the occurring crop classes.

On the one hand the method should be objective and reliable leading to an overall classification of the Seville site. On the other hand it should be transferable and operational within the MARS project. Following this considerations a processing chain consisting of four hierarchical levels has been developed.

Within level one the introduction of expert knowledge and ancillary data allows a reduction of spatial, spectral, and temporal dimensions of the dataset by masking and stratification.

Within level two the intrinsic information content of the remaining dataset (respectively single stratum) is isolated by means of a statistical sound algorithm like ISODATA clustering. This process is supported by ancillary information regarding the potential number of clusters and the number of iterations during clustering.

Within level three the final land use/crop classes are identified through deductive interpretation of the ERS SAR clusters using simultaneously general knowledge regarding agriculture (main crop types present on the site, crop calendars, tillage practices), physics (interaction of electromagnetic waves and target), and statistics (examination of cluster structures). Special emphasis has to be laid on the parcel structure/ geometry within the clustered dataset. Clusters which fit together in the corresponding parcel structure are combined and assigned to one land use/crop class.

Validation is carried out within level four using ground truth data for comparison with the classification results. In case of substantial errors interpretation flow returns back to level three, where a new clustering with different parameters has to be carried out.

5. MASKING

For an accurate classification of agricultural land use considering the limited information content of single frequency spaceborne SAR data it is useful to mask non-agricultural areas like roads, urban areas, water surfaces and forests in order to avoid misclassifications. Compared to agricultural fields these areas are characterised by their permanent surface cover types, i.e. signals detected by satellite sensors over these areas are relatively stable during a year or by year to year. Thus, reliable masks of non-agricultural land use can be created from optical satellite data already available for the MARS sites.

• Applied method

For creation of masks of non-agricultural areas an interactive hierarchical classification approach of SPOT XS data was applied (KATTENBORN et al., 1996). A separate evaluation for each theme - roads, sealed areas, forest - was performed. In a first step so called "rough masks" are created. Supported by thematic maps clearly recognisable areas of the desired classes are delineated on the screen. In a second step for every mask type a principal component analysis (PCA) is carried out using these rough masks as reference for the statistical input. In a third step for each type of mask an unsupervised cluster analysis is performed using the transformed SPOT XS bands followed by an selective assignment of resulting clusters to the respective theme. In a final step all derived masks are combined to a single mask of „non-agricultural use".

6. STRATIFICATION

On the base of a signature analysis using MARS ground truth and corresponding SPOT reflectivities or ERS-1 SAR backscatter profiles no indications for a possible stratification of the Seville site have been recognisable., since no systematic variations of signatures with respect to the geographic diversity of the site could be found. Contrary to this visual interpretation of the available satellite images (SPOT XS and PAN, ERS-1 SAR) allows a clearly recognisable differentiation of photo-morphologic units of the Seville site.

This technique was used to prepare the available information (macro-structural units, MADUE&NTILDEO, (1991)) by correcting and homogenising it for further use in ERS-1 SAR data evaluation. Important criteria for the stratification have been differences in site condition, topography, parcel structure, irrigation, number of harvests per year, and the regional agricultural practice. For practical reasons it was necessary to restrict the number of strata, but still to take into account essential differences of their agricultural use. With respect to these criteria five strata (Fig. 1) of the Seville testsite have been defined. However, the 17 MARS segments foreseen for the validation of the adopted classification approach are distributed non-uniformly over the strata; strata two and four contain each six segments, stratum five still four, stratum one only one; for stratum three ground truth is completely missing.

7. CLASSIFICATION RESULTS

The classification approach has been tested with all strata of the Seville site. Using „stratum 1" the single steps of the method are elaborated including a thorough analysis of results based on the available ground truth.

Results Stratum 1

• Description of stratum 1

Stratum 1 encloses the flat marshland of the Marismas (MADUE&NTILDEO 1991). The area is intensively cultivated, because of sufficient water supply and good soil fertility. A high proportion of rice is grown there but also winter cereals, maize, sunflower, and other summer crops (e.g. cotton) are cultivated. In most of the ERS-1 images parcel structure is discernible. This clearly recognisable geometric pattern results from dams, irrigation canals, and country lanes between the parcels. For the only existing MARS segment (No. 11) in this stratum ground survey delivered rice, maize, sunflower, water and urban areas as land use /crop types.

• Clustering

According to crop calendars many crops are already harvested half-way through the year. On these areas mostly fallow can be found after harvesting. In some areas probably the following crop was already sown respectively field preparation was done. The corresponding signals acquired with late year satellite measurements are therefore superimposed on the desirable information of spring/summer acquisitions leading to confusions during the clustering process. Thus, cluster analysis was carried out using only ERS-1 SAR data of 22.04.92, 27.05.92, 01.07.92, and 05.08.92.

According to the available a priori information five main crop types have to be expected for the stratum, namely rice, maize, winter cereals, sunflower, and other summer crops. Thereby summer crops have been considered as collective class, where no further discrimination is possible. In order to consider the backscatter variations within single crop classes due to different environmental conditions and locally varying agricultural management a sufficient number of initial clusters has to be provided to obtain homogeneous clusters with a reasonable low standard deviation. Thus, a maximum potential number of clusters of 13 was used.

• Recombination of clusters to crop classes

For the recombination and assignment of clusters to crop classes different levels and types of information have been used simultaneously. Only clusters are combined, which fit into parcel structure, i.e. which „fill up" single parcels. It has to be pointed out, that in practice smaller parts of parcel structure cannot be fully completed this way due to inhomogeneities of cluster results. Assignment of clusters to crop types was done using a priori knowledge concerning the occurring crop types and related cultivation practice , and by interpretation of backscatter signatures and statistics of the clusters. Thus, 12 resulting clusters have been finally assigned to land use classes as listed in Tab. 2.

Tab. 2: Cluster assignment to crop types and corresponding areas for stratum 1 of the Seville site.

• Spectral separability of clusters

Using the Bhattacharrya distance (RICHARDS, J. A., 1986) a separability test based on the backscatter signatures of the resulting clusters was carried out in order to assess, how far a discrimination of assigned clusters respectively crop types is possible using only their backscatter behaviour (Tab. 3). Values below 1.0 indicate a poor, between 1.0 and 1.9 a medium, and above 1.9 a good separability. In no case a poor separability was obtained. Surprisingly different clusters assigned as rice have a medium separability to other crop types but a good separability to each other (not listed in Tab. 3). By visual interpretation using parcel structure these clusters could be assigned as rice despite the more pronounced backscatter similarity to other crop types. Thus, even if separability of backscatter signatures of clusters is limited, they can reasonably be assigned using parcel structure and expert knowledge.

Tab. 3: Mean of signature separabilities for assigned clusters respectively crop types of one stratum of the Seville site estimated as Bhattacharrya distance.

• Visual inspection

Visual inspection of ERS-1 SAR classification results for the MARS segments is exemplarily shown with segment 11 in stratum 1. Every quarter of Fig. 2 shows the same subframe of the Seville site (4.6 km 4.6 km) covering MARS segment 11 and the surrounding area.

Upper left quarter depicts the five land use classes available as ground truth within MARS segment 11.

Upper right quarter shows a histogram stretched color composite of SPOT data of 01.07.92 (RGB = band 3,2,1). Vectors of segment 11 are overlaid in yellow. With a close look to the parcels at upper right and opposite upper left corner of MARS segment 11 a pronounced difference in color can be distinguished. According to MARS ground truth both parcels are mapped as maize fields. The difference in color visible in the SPOT data therefore indicate a fault in ground truth stated also by the ERS-1 SAR data classification result (lower right quarter of Fig. 2).

Lower left quarter shows the ERS-1 SAR acquisition of 01.07.92, i.e. of the same date as the SPOT data. Again vectors of segment 11 are overlaid in yellow. Flooded areas and water surfaces appear dark in this data, because of the low ERS SAR backscatter of water. From these areas the Guadalquivir can easily be recognised, others are visually identified as rice fields.

Lower right quarter shows the classification result obtained with cluster analysis of four different ERS-1 SAR images (22.04., 27.05., 01.07., 05.08.92). Legend below indicates the assigned crop types. Because of missing round truth e.g. for cotton is was not possible to further differentiate summer crop types.

• Comparison with ground truth:

Taking into account the above discussed confusion of rice and maize in MARS ground truth 830 pixels of the ground truth for rice comprising 844 pixels, i.e. 98.3%, have been correctly identified. In comparison 99 pixels of the ground truth of maize (reduced to 120 pixels), i.e. 82.5 % are actually classified as maize. From the 27 ground truth ground truth pixels of sunflower 9, i.e. 33.0 % have been correctly identified. In the case of smaller ground truth polygons like sunflower confusions, related to inaccuracies of geocoding cause major effects on the accuracy of the classification. It has to be taken also into consideration, that with the placement of segment 11 large areas of rice fields are covered. However, sunflower or maize fields are only slightly touched. Therefore, statistics for this classes just reflect the situation at the boarders of the sunflower and maize fields (Fig. 2), which is characterised by effects of mixed pixels and inaccuracies of geocoding. The remaining ground truth classes have been considered being too small for a reasonable analysis of results in this way.

RESULT STRATUM 3

During the application to the different geographical and agro-ecological conditions of the remaining strata emphasis was put on a technically understandable transfer and test of the approach. It turned out, that in certain cases the assignment of cluster results was limited to aggregated classes of crop types with the respect to restrictions of the available ERS-1 SAR time series (Tab. 1, missing data from Jan.-Mar. 1992). Due to the above discussed lack of representative ground truth a validation was performed using the available SPOT time series as shown with the example of stratum 3 in the following paragraph.

• Evaluation of stratum 3

Stratum 3, as it was originally defined on the base of macro-structural units corresponding to MADUE&NTILDEO (1991) is mainly composed of the valley of the Guadalimar river - a fertile plain with alluvial deposits. No MARS segment is located in this area.

Results of first experimentally performed clustering of the overall stratum 3 delivered inhomogeneous results, which had to be rejected. If resulting clusters have been recombined following the parcel structure of the northern part of the stratum it was not possible to simultaneously reach acceptable results for the southern part and vice versa. This can be related to differences in agricultural character and crop rotation practices between both parts of the stratum. Within the northern part almost no irrigation is applied. Within the southern part irrigation is applied to most of the area similar to stratum 1. Thus, further evaluation was performed separately for both parts of the stratum (see Fig. 3) and results have been combined afterwards.

Following available crop calendars and with respect to the limited size of stratum 3 for both parts of the stratum an initial number of 13 clusters has been fixed. The respectively resulting 12 clusters have been combined to 5 aggregated classes. The ERS-1 SAR classification result corresponds very close to SPOT NDVI information (23.03.92, 08.06.92, and 01.07.92) as it can be seen in Fig. 3, where aggregated classes are represented in colors similar to the visual appearance of the SPOT NDVI image. A general interpretation key for the SPOT NDVI composite is as follows. With regard to the sensitivity of the NDVI to photo-synthetic active biomass the color series of red, yellow, green, cyan, and blue enhances areas respectively parcels covered by a high amount of green biomass corresponding to the acquisition dates of the SPOT time series. White color indicates masked areas.

It is obvious, that with ISODATA clustering of ERS-1 SAR data and aggregation of clusters a result similar to the SPOT NDVI composite is obtained. Areas covered by an high amount of green biomass during spring time recognisable due to their red color in the SPOT NDVI composite (predominance of the SPOT NDVI image of 23.03.92) are mostly also identified by the ERS-1 SAR classification. Using a priori knowledge these areas can be considered to be winter cereals or also grassland. From Fig. 3 it is obvious, that in the northern part of stratum 3 more winter crops or grassland are found than in the southern part.

Areas covered by a high amount of green biomass during summer time recognisable due to their cyan color in the SPOT NDVI composite (predominance of the SPOT NDVI images of 08.06.92 and 01.07.92) and mainly occurring in the southern part of the stratum can be interpreted as summer crops.

Also with respect to parcel structure in the SPOT NDVI composite as well as in ERS-1 SAR classification results the differences in cultivation practice as applied in the northern and the southern part of stratum 3 are evident. More detailed descriptions of results for stratum 1 and 3 as well as evaluation and results for the remaining strata of the Seville site can be found in KATTENBORN, G., et al. (1996).

8. CONCLUSION

A time series of nine ERS-1 SAR images of the MARS sample site Seville, four SPOT XS images and ground surveys of 17 MARS segments all from 1992 have been available for the study. During preliminary investigations it was figured out, that significance of supervised classification results is limited. First available input data (MARS segment ground truth) are not representative for the overall testsite respectively for single strata. Second separability between the ERS backscatter signatures of ground truth polygons is poor. Additionally histograms of most of the backscatter signatures showed up multi-modal distributions.

Thus, a newly developed robust approach, also transferable to other MARS sites was developed.

In order to exclude areas with permanent use from ERS SAR classification preliminary masks derived from thematic maps were refined by multi-temporal, unsupervised classification of optical satellite data in order to create masks of main permanent land use classes (roads, water surfaces, urban areas, forest). Stratification was performed using visual interpretation of SPOT data with respect to agro-ecological phenomena.

For the following classification ERS-1 SAR data are statistically segmented. Precise knowledge about the local agricultural practice, i.e. especially the crop rotation scheme, is necessary in order to identify all useful acquisition dates of the available ERS-1 SAR time series. With respect to crop calendars more ERS-1 SAR acquisitions than available for this study are required especially in the early growing season.

Recombination of clusters to classes is performed using ancillary knowledge (local agricultural practice, former agricultural statistics) and parcel structure (road and country lane network or a cadastral data base).

The method was developed and analysed in detail with the example of a single stratum delivering excellent results. The transfer to the other strata of the Seville site proved functionality of the method and delivered reasonable results just limited to a preliminary level by the lack of sufficient a priori information. The method proved robust against effects superimposing the desirable information (local agricultural management, agro-ecological conditions and related phenological differences, geographical diversity). It was shown, that ERS SAR data supply information similar to SPOT NDVI images, thus giving significant indications, that ERS SAR data could be used as complement or even as substitute to optical satellite data.

In terms of the operational needs of the MARS project the potential of the approach seems to be high. With regard to the processing of ERS-1 SAR data necessary algorithms are available as standard software. With regard to masking and stratification site-dependent effort is required. If once GIS data for masking and stratification have been compiled and implemented only periodical updates are necessary. In future additional data sources can be expected from administrational data bases in most European countries (e.g. ATKIS in Germany). The recombination and assignment of clusters requires interactive control, but can also be supported by GIS information up to a certain extent. For the rationalisation of this task expert systems will possibly play an important role in future. In conclusion a comprehensive and robust concept for the operational use of the continuous ERS SAR data flow for agricultural statistics was developed and its plausibility demonstrated.

ACKNOWLEDGEMENTS

This study was funded by JRC/Ispra under contract No. 10161-94-04 F1ED ISP D. We would like to thank Dr. Hugo De Groof, Dr. Hans Kohl and Dr. Edmond Nezry for their co-operation and effective support.

REFERENCES

DOBSON, M. C., PIERCE, L., SARABANDI, K., ULABY, F. T. and SHARIK, T. (1992): Preliminary analysis of ERS-1 SAR for forest ecosystem studies. IEEE Trans.Geosci. Rem. Sens., Vol.30, pp. 203-211.

HARTL, P., G&UUMLTH, S., KLAEDTKE, H.-G., PALUBINSKAS, G. , REICH, M. and W&OUMLRZ, K. (1995): Application of multi-temporal ERS-1 SAR data: Some results of the PASTA project. Interim Report to ESA, Institute of Navigation, University of Stuttgart, Stuttgart, 27 pp.

KATTENBORN, G., KLAEDTKE, H.-G., G&UUMLTH, S. and REICH, M. (1996): Potential of ERS-1 SAR for agricultural statistics. Final Report, JRC/CEC Contract no. 10161-94-04 F1ED ISP D. 143p.

KOHL, H. G., NEZRY, E. and DE GROOF, H. (1994): Crop acreage estimation with ERS-1 PRI images. Earth Observation Quarterly, 46, December 1994, pp. 6-9.

LAUR, H. (1992): ERS-1 SAR calibration: derivation of the backscattering coefficients in ERS-1 SAR PRI images. Technical Note, ESA/ESRIN, 17 pp.

LOPES, A., NEZRY, E., TOUZI, R. and LAUR, H. (1993): Structure detection and statistical adaptive speckle filtering in SAR images. Int. J. Remote Sensing, Vol.14, No. 9, pp.1735-1758.

MADUE&NTILDEO, J.M.M. (1991): Capacidad de uso y erosion de suelos. Una aproximación a la evaluación de tierras in Andalucía. Junta de Andalucía, Agencia de Medio Ambiente, 446 pp.

NEZRY, E., GENOVESE, G., SOLAAS, G. A., REMONDIERE, S, AND KATTENBORN, G. (1995): Early crop identification and area estimation in Europe using ERS winter images. Accepted by IEEE Trans. Geosci. Rem. Sens.

RICHARDS, J. A. (1986): Remote sensing digital image analysis. Springer Verlag, Berlin, Heidelberg, New York, London, Paris, Tokyo, pp. 206 - 225.

Fig. 1: Panchromatic SPOT scene from 28.12.92 of the Seville site overlaid with vectors of area under investigation (white), macro-structural units (corr. to MADUE&NTILDEO, 1991, green, small) defined strata S1...S5 (green, broad) and MARS segments 1...17 (red).

Fig. 2: Satellite data and ERS-1 SAR classification result for a subframe of the stratum around MARS segment 11 of the Seville site. UL: Ground truth bitmaps; UR: SPOT false color composite of 01.07.92; LL: E