Abstract

A multi-spectral SPOT image, polarimetric airborne SAR data as well as satellite based C-band SAR data have been used to perform classification of agricultural fields and areas occupied by forest and take. Conventional Maximum Likelihood classification has been compared with classification incorporating a Gaussian mixture class model, as well as an algorithm based on multi-resolution structured data and sequential MAP (SMAP). The classification accuracies found were generally high, using combinations of sensors. It is found that multi polarization data gives invaluable information to be used in a classification scheme, a feature that can be exploited in future satellite sen sors, like for instance ASAR on board ENVISAT. The Gaussian mixture class model performed only slightly better than the conventional maximum likelihood algorithm, whereas the SMAP algorithm improved the classification results.

Keywords

C'rap classification. Remote Sensing, Synthetic Aparturc Radar, combining
data, contextual classification

Inge Sandholt, Institute of Geograph\, Universitv of Copenhagen,
Ostervoldgade 10, DK-1350 København K, Denmark,
e-mail is@geogr.ku.dk

(ieocrafisk Tidsskrift. Danish Journal of (ieocranhv 101 ' 21-12

Since the first SAR. data became available, several attempts to use the data for land use classification have been made. Dobson et al. (1995) compare several SAR classifiers, reporting a wide range of accuracies, spanning from 75% to more than 98%. The studies base primarily the classification on C-band, L-band polarimetric data, or on combinations hereof. Generally, the highest accuracies are found for multi-frequency combinations, and for studies using few classes (three or less). The potential of discriminating crop types using microwave data have been reported by Wegmuller (1993) and Skriver et al. (1999). Wegmuller (1993) concludes that different microwave parameters can be used to distinguish different crops, but maybe more important, the temporal evolution in radar returns for individual crops can be exploited successfully via a multi-temporal data set. This is in agreement with the results found by for instance Bouman and Hoekman (1993), Sandholt et al. (1995), and Skriver et al. (1999). Another study (Anys & He, 1995) reports from a C-band multi-polarization data set, that using supervised classification methods, the accuracy can be improved considerably including textural features (1^st, 2^nd and 3^rd order histograms). Texture measures are difficult to include in multi-temporal data sets though, due to the need

tø co-rca!stcr !rnnlv!na rcscimn!ina of the !måaes which
*V^*LJIV*) »».»J.^.^»»»^ .•*.•.. t-.,. » »|^ ... .^ v^« ...x. ... . »^•w u 5
may alter the textural information in the images. Textural
information should be extracted prior to any resampling of
data. Very high classification accuracies have been reported
in Chen et al. (1996) using a neural network to classify an
agricultural area into 12 classes basedon multi-frequency
polarimetric SAR data. On the other hand, single polarized
C-band SAR data alone have not proved well suited for
classification purposes, see for instance Durand et al.
(1987). Schotten et al. (1995) base a classification of agricultural
land on a multi-temporal data set of ERS-1 images.
The area contains a very high number of agricultural fields
(980), allowing a field based classification to be done. The
overall classification result obtained is 80%.

Several studies have shown that SAR data may provide information on structural features of the surface complementary to the spectral information in optical data, for instance Dobson et al. (1995), Sandholt et al. (1995), Solberg et al. (1994), and Morgan et al. (1992). It has thus been shown, that the discriminating power of SAR images is very much improved when they are used in combination with optical data. Solberg et al. (1996) have compiled a brief but useful review of multi-source classification.

In this paper, classification of agricultural land by means of remote sensing is addressed. The primary goal is to show how multi-source remotely sensed data can be successfully combined to perform classification of agricultural land with respect to crop type, and secondly to test the performance of existing algorithms based on Bayesian classification techniques. Some of the methods belong to the group of contextual classification, and exploit the ideas behind multi-resolution data representation. The data set is quite unique, in the respect that it comprises temporal C-band SAR data with different spatial resolutions, and polarimetric SAR data, together with an optical satellite image. Part of the study will show the effects of using polarimetric SAR data from an airborne SAR in combination with optical data, a very relevant approach to prepare for EN VIS AT data, or for ERS data in combination with RADARSAT to obtain multipolarization data. The paper is organized as follows. First a brief introduction to the applied methods is given. Some remarks are made on contextual classification including Markov Random Fields, on which one of the applied algorithms is based. The test site and data are presented , followed by results and discussion. Finally, the overall results are summarized in the conclusion.

Classification algorithms

Discriminant analysis

In land use classification discriminant analysis can be applied when the class identities of some samples (or pixels) are known a priori. Two sets of multivariate observations for which the class membership is known are thus chosen, a training set, and a test set, allowing an independent evaluation of classification results. In discriminant analysis, the class membership of the test set is determined based on the statistics of the training set. We will here consider classical discriminant analysis, a technique described in most textbooks on multivariate analysis, for instance Everitt & Dunn (1991), Krzanowski (1988) or Mardia et al. (1979). A comprehensive exposition of discriminant analysis related to pattern recognition can be found in McLachlan (1992). The classification performed is a conventional per pixel maximum likelihood classification.

The gaussian mixture class

The objective of mixture classes is to improve classification performance by modelling each class as a pro-babilistic mixture of a variety of subclasses. The approach is thus useful when each class contains pixels with a variety of distinct spectral characteristics, like for instance forest with different tree species, or urban areas with streets and buildings. For agricultural crops, where classes are likely to be more homogeneous, the Gaussian mixture class is useful when pixels in each class consist of different sorts of crops. For each class a clustering of subclasses is performed using the method described below. The mixture density for k classes in the multivariate case is given by:

d)

Typically, the number of subclasses in each group are not known a priori. To determine the number of subclasses Rissanen's minimum description length (MDL) has been applied (Rissanen, 1983). Another possibility would have been to use Akaike and Schwarz's information criteria (Schwarz, 1978). Both the number of distinct subclasses in each class and the spectral mean and covariance for each subclass are estimated by the algorithm. More on how this is handled can be found in Bouman and Shapiro (1996).

Contextual classification

In applying classical discriminant analysis, no attention is paid to the spatial dependency between adjacent pixels. This may result in a spotty appearance of the resulting classification image, and also that important information contained in the image is neglected. Therefore attempts have been made to exploit the spatial information. These methods are somewhere in the area between segmentation methods and classification methods, and are denoted contextual classification, not to be confused with methods taking the texture of the images into account. Contextual classification deals with the relationship between adjacent pixels, whereas textural classification addresses the composition of the neighbourhood. The latter can be done simply by including texture information in the feature vector. Examples of including texture parameters of SAR data in image classification can be found in for instance Solberg et al. (1994), Morgan et al. (1992), and Anys & He (1995). Context can be considered along three different dimensions: the spectral, the spatial and the temporal dimension (Solberg et al., 1996). Markov Random Fields (MRF) provide a methodological framework to exploit spatial information in images, and several studies have been reported. Therrien et al. (1996) utilizes a model for texture based on stochastic linear filtering concepts to segment aerial photographs of a rural area. The texture model is combined with a MRFmodel

to represent the occurrence of textured regions within an image. He concludes that the MRF-model is superior to a maximum likelihood model for the same image, and that the difficult optimization problems can be overcome by applying a suboptimal procedure in a combination of the texture model and the MRF. Haslett (1985) formulated a widely used classifier based on maximum likelihood discriminant analysis using a Markovian model of spatial context. An early example of a contextual classifier, is ECHO, Extraction and Classification of Homogeneous Objects (Kettig & Landgrebe, 1976). Successful attempts have also been made to apply methods that model spatial context as well as temporal context, see for instance Jeon and Landgrcbe (1992) and Solberg et al. (1996).

Multi-scale Markov Random Fields

One of the classification algorithms used in this study, is based on a multi-scale Markov Random Field model developed by Bouman & Shapiro (1994). A Markov Random Field is a concept originating from time series analysis. In one dimension, a Markov chain can be defined as a process satisfying (Cressie, 1991):

where {Z(t):t = 0,l }isa random process in time where data are observed at the times 0,1 „..n. Equation 2 expresses the lack of memory property that the conditional probabilities for the present given the past is determined only by the most recent observation. In time series analysis, the definition of the past is straightforward, but when the model is extended to two dimensions several possibilities are available. In the spatial case, the concept of neighbourhood is introduced, often as a set of four or eight nearest pixels. The appealing feature of MRF, is that they only require the specification of spatially local interactions using a set of local parameters (Bouman & Shapiro, 1994).

The algorithm developed by Bouman and Shapiro (1994) in short uses a multi-scale random field (MSRF), that is a Markov Random Field in scale, and the estimation is done by a sequential maximum a posteriori (SMAP) algorithm. The MSRF is composed of a series of random fields progressing from coarse to fine scale. The Markov property assumed is that each field depends on the previous coarser field only, and the Markov chain is thus in scale . Further it is assumed that points in each field are conditionally independent given their coarser scale neighbours. The SMAP estimation method consists in minimizing the expected size of the largest misclassified region, in a manner such that progressively

larger costs are assigned to errors at coarser scales.The algorithm has been described in detail in Bouman and Shapiro (1994) and in Bouman & Shapiro (1996). The actual implementation used here is described in Grass4.l (1993). The SMAP algorithm has been applied on SAR data from Sweden by Michelson et al. (2000).

Experimental data

The study area is located near Research Centre Foulum in the central part of Denmark, and comprises agricultural land and forest belonging to Tjele Estate - where a 5.12x5.12 km² subset has been selected.

The classifications are based on a subset of the DANMAC 94 data set consisting of ERS-1 SAR data and a SPOT image (Thomsen et al., 1994). In addition, data from the Danish airborne SAR, EMISAR are used. The EMISAR is the result of a research and development project initiated in 1986 at the Department of Electromagnetic Systems (EMI) of the Technical University of Denmark, and it is a fully polarimetric and interferometric L-and C-band SAR (Madsen et al., 1991; Christensen et al., 1998). For this study, only the C-band SAR has been applied. The FMISAR is used for scientific experiments conducted by the Danish Center for

Remote Sensing (OCRS) which was established in 1994 at EMI by the Danish National Research Foundation. The SAR system is flown on a Royal Danish Air Force Gulfstream G-3 aircraft. The SAR system is normally operated from an altitude of approximately 12,500 m, the spatial resolution is 2 m by 2 m, the ground range swath is approximately 12 km and typical incidence angles range from 35 deg. to 60 deg. The processed data from this system are fully calibrated by using an advanced internal calibration system. The multi-temporal image data consist of three ERS-1 SAR images, two polarimetric airborne C-band images from the EMISAR. and the SPOT image.

For the EMISAR data only the VV, HH and HV polarized intensities were used. The most important specifications of the sensors and platforms are summarized in Tables 2, 3 and 4.

The timing of image acquisitions is nearly optimal with respect to crop evolution, only an additional early acquisition from May would have been desirable, see for instance Wegmuller (1993) or Sandholt et al. (1995) for exploitation of the temporal evolution in backscatter measured by SAR during the entire growing season. The classification algorithms require Gaussian distributions, and the SAR intensity data cannot be used directly, as the pdf for intensity data can be approximated by a Gamma distribution. One way of working around this is to logtransform the SAR images (Sandholt et al., 1995) prior to further processing and as such a Gaussian assumption is plausible. This approach has

been used here. Another solution would have involved reformulation of the classifier taking the SAR statistics into account, as for instance have been done by Solberg et al. (1996).

All images have been geometrically corrected to UTM, and resampled to 5x5 m2m² pixels in order to be able to superimpose the images. The resampling involved a smoothing of the EMISAR data with a 9x9 weighted mean filter. A number of image and band combinations have been used in the classifications, as listed in Table 5.

A number of fields have been selected for reference and test of results. Six different crops are grown on the selected fields, and together with samples from Tjele Lake and Tjele forest, they comprise the eight cover classes in the study. The set of fields has been divided into a training set and a test set as indicated in Figure 3 and Table 6.

The wheat fields contain four different varieties, and the fields have been divided such that the training set consists of the varieties hereward, pepital and husar and the test set consists of the varieties hereward, husar and ritmo. This may result in lower classification performance for wheat than may have been found if pepital and ritmo had been excluded from the data set. However, to keep as many observations in the data set as possible, all four varieties have been used.

Text confines on page 26

Results and discussion

Classical discriminant analysis

The first task is to perform conventional per pixel maximum likelihood classification. Table 7 shows the classification accuracy for each class in the test set, and the average accuracy for the combinations.

The average accuracy for the ERS image combination (40.0%) is clearly not acceptable. A similar study (Sandholt et al., 1995) with a comparable number of classes and number of ERS-1 images found an average accuracy 51.7% for ERS-1 images. A study using eight ERS-1 images based on per field classification, found accuracies up to 80% (Schotten et al., 1995). The higher accuracies in the latter study may be due to the per field approach, which to some extend also minimizes the effects of speckle, and to the higher number of images available. Average classification accuracy based on the two EMISAR images (80.6%) is comparable to what is found for the multi-spectral SPOT image (83.8%), and only a minor improvement is seen when the two SAR sensors are combined (81.1%). However, the classification based on SPOT has its lowest performance for wheat (23.9%), probably due to the fact that the wheat fields in fact encompass three different varieties of wheat. It should be noted that, although not satisfactory, the SPOT based crop classification accuracy is higher than what could have been expected for a single scene, see for instance Fog et al. (1993) or Sandholt et al. (1995). The most severe problem for the EMI combination, is found for barley with a classification accuracy of 54.0%. 38.0% of the barley pixels in the test set are classified as grass.

Gaussian mixture class classification

The expected improvement of the classification accuracies
using Gaussian mixture class classification was not found
(TableS).

The mixture class model allows each class to consist of several subclasses, and for not entirely homogeneous classes like forest an improvement would have been expected. However, the accuracy improvement of the accu

racy is small, and for some classes the accuracy is compara
tr> nr (^\/<=>n Ir\\\/p>r f Via n fr\r tliA m-jvimum

likelihood classification (Table 7). The fact that the selected fields are rather large and homogeneous, leaves the only slight improvement to be found for the more complex classes, forest and wheat, the latter comprising four varieties of wheat. The average accuracy for wheat using the SPOT data alone is 4% higher when using the mixture class model, but still the accuracy is generally low for wheat, probably due to the fact, that the training set did not comprise all the wheat varieties.

Turning to a comparison of the chosen sensor combinations, the best average performance is found for the combination comprising the total data set. The average accuracy is 94.5, and except for wheat, the accuracy for each crop type is higher than 90%, and for rye, oat, oil seed rape and lake the accuracies are even equal to or higher than 98.0%. The average classification accuracy for the combination SPOT EMI is comparable to the classification accuracy for the total data set, showing the little effect of including ERS-1 data when EMISAR data is already available. More on the different sensor combinations can be found in the next section.

Multi-scale contextual classification

In Table 9 the multi-scale classification accuracies are shown. The Gaussian mixture class model is included in the multi-scale algorithm. Comparison of Tables 8 and 9 clearly expose how the multi-scale contextual classification algorithm outperforms the previously used algorithms. The ;mr>rr.w^tY-i^rit '< r. n>^rf^rT-n o r-,,>^ i, L.r.ru.l f,xr-11-,_{t >} rr,,,_t-» nr^i-i/ UllplVrrWlllWllt ... KX,*IVSIIIIUIIVVK^{X,*IVSIIIIUIIVV} „, .H,&^l .W. 111 V 111WJI llWUjr images with the highest resolution. For the SPOT image with 20 m pixel spacing, the relatively coarse resolution (at least in comparison to EMIS AR data) has the effect that the hierarchical data structure is not fully exploited. Classification based on the speckly SAR images in contrast, takes advantage of the simple segmentation inherent in the algorithm.

The confusion matrices for classification based on the SPOT image and EMISAR images are shown in Tables 10 and 11 and the confusion matrix for classification based on the total data set is shown in Table 12.

Table 10 clearly shows how the SPOT-based classification is not able to discriminate the wheat pixels, 63.5% are classified as rye. It also fails to classify the grass pixels, only 6.44 % are correctly classified as grass. Instead the grass pixels are classified as rape. In contrast the EMISAR based classification confuses barley and grass, the two crops having a similar structure in the early growing stages. When the two data types are combined, the confusion between the mentioned groups are considerably lower, as can be seen in Table 12, and the discrimination between the other classes is improved as well. The highest average classification accuracies are found for the total data set and for the SPOT EMI combination (95.9% and 95.5% respectively), as was the case in per pixel Gaussian mixture class classification, only the accuracies here are improved considerably. The

classified image with the best average accuracy can be seen
in Figure 4.

The results are in agreement with what has been found earlier in a comparison between SMAP and Maximum Likelihood classification of McCauley and Engel (1995), who based on airborne scanner data for eight classes, found SMAP to perform better than conventional maximum likelihood classification. Solberg et al. (1996) report overall classification accuracies for a Landsat TM image and a multi-polarization C-band SAR image from the MAESTRO experiment acquired seven days apart. They achieved 71.3% classification accuracy for their simple fusion model, and 73.0% for their Markov random field model based classification in a 7 classes classification problem. The accuracies found in the present study are significantly higher, if we compare with the SPOT PM6 combination in Table 9. For comparison of the two studies, we should be aware, that the overall accuracy (number of correctly classi

fied pixels in relation to the total test set) typically is less than the average accuracy (average of number of correctly classified pixels in relation to total number of pixels within each class). In the same study, by Solberg et al. (1996) accuracies of combining Landsat TM with ERS-1 SAR (5 classes) for a diiiereni iesi sue were iouud io be unproved when the Markov random field model was applied. The improvements found by applying the multi-scale classification model in the present study are considerably higher, but again the two studies are not directly comparable; for instance, the reference models used here (ML-classification and Gaussian mixture class classification) are simpler than the one applied in the study reported by Solberg et al. (1996), and the timing of the image acquisitions is not comparable. In addition, it is well known, that Landsat TM due to its wavelengths further up in near infra red band is better suited for vegetation classification than SPOT. Typically users will not have the airborne SAR data available, often only satellite data can be included in the feature vectors.

/~< ~ii *~ii:«.~ c A o „^„,,^„,, „-^ i:,v,;»^^i <-/-> c-',r,r,\o
v_.ui i ciiuy .>aiGiiiic orviv n^naui 3 aiv_ iiiimvu LW wuipm jni£,iv.

polarized data. The multi-polarized data are expected to provide valuable information easily exploited in classification algorithms, and the above results makes that argument even stronger. Therefore two vertically polarized EMISAR images have been combined with the SPOT image._lt should be noted that the vertically polarized EMISAR data is not a simulation of ERS-1 data with a higher spatial resolution due to the fact that the incidence angles for ERS-1 and

EMISAR arc very difterent. and thus the backscattcrcoeilicients åre different. However, the combination enables us to analyse how inclusion of multi-polarized data improves the classification accuracy by comparison with the single polarized data. The results are found in Table 9. and it is interesting to note, that the accuracies for the SPOT EMlcvv combination in general match the accuracies for SPOT ERS apart from wheat and grass. The average accuracy is considerably higher for the SPOT EMI combination (95.59É versus 89.0%). leading to the conclusion that multi-polarized SAR data is a valuable data source in classification of crop types. The study done by Dobson et al. (1996) based on ERS-1 (C-band, vv-polarization) and JERS-1 data (L-band. hh-polarization), reports similar improvements in classification accuracies when multi-polarization data are used. However, the composite images used in that study, are multi-frequency as well, and a major part of the improvement could be found here.

Conclusions

A unique data set was used to perform classification of agricultural
land with respect to six different crop types and areas
occupied by forest and lake. The performance of a clas

sification algorithm based on Gaussian mixture class, and of a multi-scale classification algorithm were tested. Also different image combinations were tested. The best combination of images was the one comprising all of the available data, but only little difference in overall classification accuracies was found between the combination consisting of the SPOT image and the airborne multi-polarized SAR data (EMISAR). In addition to that, it was found, that multi-polarized data gives invaluable information to be used in a classification scheme, a feature that can be used in future satellite based sensors like ASAR on board ENVISAT. In fact, classification accuracies based on multi-temporal, polarimetric SAR data, were higher than for a single SPOT

scene. That may have important implications for classification set ups in operational applications, when ENVISAT is launched, due the all weather capabilities" of SAR sensors, which ensures a multi-temporal data set. However, one has sensors and airborne sensors due to the differences in incidence angles and spatial resolution. However, the increase in classification performance of polarimetrie data is signifi

cant.

Thc Gaussian mixture class model was found to improve
the scores for forest and wheat, the latter which comprised
lour varieties, but generally, the mixture class model does

not improve classification accuracies for agricultural fields, due to the relatively homogeneous classes, and in some instances the accuracies are even decreased. The multi-scale classification algorithm (SMAP) clearly outperformed the per pixel classification. The largest improvement was found for the relatively noisy high resolution SAR data (EMIS AR), whereas the improvement for the coarser scale, less noisy SPOT data was much less.

Using the SMAP algorithm on the total data set, an average classification accuracy of 95.9% was achieved, which is considered to be satisfactory, specifically taking into account, that neither multi-spectral SPOT data nor C-band SAR data are measured at optimal wavelengths with respect to vegetation monitoring. The result is expected to be improved if Landsat data is included, taking advantage of the near infrared band in Landsat data.

Acknowledgements

The work was done while the author was with the Department of Mathematics and Physics at the Agricultural Universitv of Denmark. It has been funded bv Dina. the Danish Informatics Network in the Agricultural Sciences, and by the Danish Space Board via UANMAC, the Danish Multisensor Airborne Campaign. The EMISAR data were acquired and processed by the Danish Center for Remote Sensing. Development of the EMISAR has been co-sponsored by the Thomas B. Thriges Foundation, the Danish Technical Research Council (STVF), the Royal Danish Air Force, the Technical University of Denmark, and the Joint Research Centre. The data acquisitions and the data processing have been sponsored by the Danish National Research Foundation. The author wishes to thank Henning Skriver, DCRS, for usefull comments to the manuscript and help and advice.

References

Anys, H. & He, D.-C. (1995): Evaluation of textural and multipolarization
radar features for crop classification. lEEE
Transactions on Geoscience and Remote Sensing, 33(5).

Bouman, B. & Hoekman, D. (1993): Multi-temporal, multi-frequency radar measurements of agricultural crops during the Agriscatt-88 campaign in The Netherlands. International Journal of Remote Sensing, 14(8): 1595-1614.

Bouman,C. A. & Shapiro, M. (1994): Amultiscalerandom

field model for bayesian image segmentation. lEEE
Transactions on Image Processing, 3(2): 162-177.

Bouman, C. A. & Shapiro, M. (1996): SMAP image segmentation software package. Users manual, http:// dynamo.ecn.purdue.edu/ bouman/software/segmentation_and_clustering/.

Chen, K. S., Huand, W. P., Tsay, D. H. & Amar, F. (1996): Classification of multifrequency polarimetric sar imagery using a dynamic learning neural network. lEEE Transactions on Geoscience and Remote Sensing, 34(3): 814-820.

Christensen, E., Skou, N., Dall, J., Woelders, K., Jorgensen, J., Granholm, J. & Madsen, S. (1998): EMISAR: An absolutely calibrated polarimetric L- and C-band SAR. lEEE Transaction Geoscience and Remote Sensing, 36(6):1852-1865.

Cressie, N. (1991): Statistics for spatial data. John Wiley
and Sons, New York.

Dobson, M., Pierce, L. E. & Ulaby, F. T. (1996): Knowledge-based land cover classification using ERS-1/JERS-1 SAR composites. lEEE Transactions on Geoscience and Remote Sensing, 34(l):83

Dobson. M. C.. Ulahv. F. T. &. Pierce. L. F. M 995 V Landcover classification and estimation of terrain attributes using synthetic aperture radar. Remote Sensing of Environment, 51:199-214.

Durand, J. M., Gimonet, B. J. & Perbos, J. R. (1987): SAR
data filtering for classification. lEEE Transactions on
Geoscience and Remote Sensing, 25(5):629-637.

Everitt, B. S. & Dunn, G. (1991): Applied Multivariate Data
Analysis. Edward Arnold, London.

Fog,B.,Poulsen,J. N.,Sandholt,l.,Stjernholm,M.& Skriver, H. (1993): Monitoring land cover and crop types in Denmark using ERS-1 SAR and optical satellite images. Pp. 71 -80 in Proc. Int. Symp. From Optics to Radar, SPOT and ERS applications, Paris May 5-9 1993, number 312 in CNES.

Grass4.l (1993): Grass 4.1 User's Reference Manual.
USACERL, Champaign, Illinois.

Haslett, J. (1985): Maximum likelihood discriminant analysis
on the plane using a Markovian model of spatial context.
Pattern Recognition, 18(3):287

Horgan, G. W., Glasbey, C., J.N., Cuevas Gozalo, S. L. S. & Alonso, F. G. (1992): Land-use classification in central Spain using SIR-A and MSS imagery. International Journal of Remote Sensing, 13(15):2839

Jeon, B. & Landgrebe, D. A. (1992): Classification with
spatio-temporal interpixel class dependency contexts.
lEEE Transactions on Geoscience and Remote Sensing,

30(4):663-672.

Kettig, R. & Landgrebe, D. A. (1976): Classification of multispectral image data by extraction and classification of homogeneous objects. lEEE Transactions on Geoscience Electronics, 14(1): 19-26.

Krzanowski, W. J. (1988): Principles of multivariate analysis:
a users perspective. Oxford Statistical Science Series.
Clarendon Press.

Madsen, S., Christensen, E., Skou, N. & Dall, J. (1991): The danish sar system - design and initial tests. lEEE Transactions on Geoscience and Remote Sensing, 29:417-426.

Mardia, K. V., Kent, J. T. & Bibby, J. M. (1979):
Multivariate analysis. Academic Press

McCauley, J. D. & Engel, B. A. (1995): Comparison of scene segmentations: SMAP, ECHO and maximum likelihood. lEEE Transactions on Geoscience and Remote Sensing, 33(6): 1313-1316.

McLachlan, G. J. (1992): Discriminant analysis and statistical pattern recognition. Wileys series in Probability and Mathematical Statistics. John Wiley and Sons, New York.

Michelson, D., Liljeberg, B. & Pilesjo, P. (2000): Comparison of algorithms for classifying Swedish landcover using landsat tm and ers-1 sar data. Remote Sens. Environ.

Rissanen, J. (1983): A universal prior for integers and estimation
by minimin description length. Annals of Statistics,
11(2):4171(2):417-

Sandholt, L, Fog, 8., Poulsen, J. N., Stjernholm, M. &
Skriver, H. (1995): Classification of agricultural crops in
Denmark using ERS-1 SAR and SPOT imagery. Askne,

J., editor, Sensors and Environmental Applications of Remote Sensing, Proceedings of the 14th EARSeL Symposium, Göteborg, Sweden 6-8 June 1994, pages 37-44. Rotterdam. A.A. Balkema Publishers.

Schotten, C. G., Rooy, W. W. L. V. & Janssen, L. L. F. (1995): Assessment of the capabilities of multi-temporal ERS-1 SAR data to discriminate between agricultural crops. International Journal of Remote Sensing, 16( 14): 2619-2637.

Schwarz, G. (1978): Estimating the dimension of a model.
The Annals of Statistics, 6(2):461 -464.

Skriver, H., Svendsen, M. T. & Thomsen, A. G. (1999): Multitemporal C- and L-band polarimetric signatures of crops. lEEE Transactions on Geoscience and Remote Sensing, 37(5):2413-2429.

Solberg, A. H. S., Jain, A. K. & Taxt, T. (1994): Multisource classification of remotely sensed data: Fusion of Landsat TM and SAR images. lEEE Transactions on Geoscience and Remote Sensing, 32(4):768-785.

Solberg, A. H. S., Taxt, T. & Jain, A. K. (1996): A Markov random field model for classification of multisource satellite imagery. lEEE Transactions on Geoscience and Remote Sensing, 34( 1): 100-113.

Therrien, C. W., Quatieri, T. F. & Dudgeon, D. E. (1986):
Statistical model-based algorithms for image analysis.
Proceedings of the lEEE, 74(4).

Thomsen, A., Fog, 8., Skriver, H. & Stjernholm, M. (1994):
The DANMAC data report 1994. Danish Institute of
Plant and Soil Science.

Wegmuller, U. (1993): Signature research for crop classification
by active and passive microwaves. International
Journal of Remote Sensing, 14(5):8711-