A NEW POLARIMETRIC CLASSIFICATION APPROACH EVALUATED FOR AGRICULTURAL CROPS

Transcription

1 A NEW POLARIMETRIC CLAIFICATION APPROACH EVALUATED FOR AGRICULTURAL CROP, ) Dirk H. Hoekman ) Wageningen University Dept. of Environmental ciences, ub-department of Water Resources Nieuwe Kanaal, 679 PA Wageningen, The Netherlands tel:+-7-9, fax:+-7-5, Dirk.Hoekman@wur.nl ) arvision BV, ABTRACT tatistical properties of the polarimetric backscatter behaviour for a single homogeneous area are described by the Wishart distribution or its marginal distributions. These distributions do not necessarily well describe the statistics for a collection of homogeneous areas of the same class because of variation in, for example, biophysical parameters. Using Kolmogorov-mirnov (K-) tests of fit it is shown that, for example, the Beta distribution is a better descriptor for the coherence magnitude, and the log-normal distribution for the backscatter level. An evaluation is given for a number of agricultural crop classes, grasslands and fruit tree plantations at the Flevoland test site, using an AirAR (C-, L- and P- band polarimetric) image of July 99. A new reversible transform of the covariance matrix into backscatter intensities will be introduced in order to describe the full polarimetric target properties in a mathematically alternative way, allowing for the development of simple, versatile and robust classifiers. Moreover, it allows for polarimetric image segmentation using conventional approaches. The effect of azimuthally asymmetric backscatter behaviour on the classification results is discussed. everal models are proposed and results are compared with results from literature for the same test site. It can be concluded that the introduced classifiers perform very well, with levels of accuracy for this test site of 9.% for C-band,.7% for L- band and 96.% for the combination of C- and L-band.. INTRODUCTION With the advent of polarimetric spaceborn AR systems like RADARAT- (C-band) and ALO PALAR (L-band) the interest for simple, robust and accurate polarimetric classification and biophysical parameter estimation algorithms for monitoring applications is increasing. In this paper a number of classification approaches applied on the Flevoland agricultural test site in The Netherlands will be discussed. In the literature [-] several results for this particular test site using data from the MAETRO- campaign can be found enabling comparison of performances to a certain extent. During the MAETRO- campaign Flevoland was imaged by NAA/JPL s C-, L- and P-band polarimetric AirAR system at 6 August 99. In [] the use of a simple hierarchical parallelepiped algorithm is discussed. The crops are grouped in 9 classes. Using 7 polarisations: HH, HV,VV, RR, RL, 5C (+5/+5 linear) and 5X (+5/-5 linear), and all three frequency bands, a classification accuracy level of 7.5% is achieved for x pixel aggregates and a level of 9.5% at a per-field basis. In [] the use of optimal polarisation selection and a wavelet-based texture feature set is discussed. For classes, including water, using three frequency bands and three synthesized optimised polarisations an accuracy of 9.% is achieved. Using the HH, HV and VV polarisations instead, the accuracy reduces to only.6%. In [] the same data set is used. Applying a dynamic learning neural network the accuracy increases to 95.%. However, for C-band only the result reduces to 67.9% and for L-band to 7.%. In [] the use of the complex Wishart distribution for the covariance matrix and the use of ML classifiers, for different polarisation combinations, are discussed. All pdf s are derived from the complex Wishart distribution under the circular Gaussian assumption for complex polarimetric data. For classes, including water, the best single band case is L-band fully polarimetric with.6%. C-band achieves 66.5%. Using all three bands a level of 9.% is reached. In this paper a polarimetric classification technique developed by the author for application in heterogeneous tropical regions [5] is tested on the Flevoland test site and some suggestions for further improvement are made. In this approach statistical properties of the polarimetric backscatter behaviour for a single homogeneous area are described by the complex Wishart distribution, like in []. However, these distributions do not necessarily well describe the statistics for a collection of homogeneous areas of the same class because of variation in, for example, biophysical parameters,

2 which often is reflected in variation in the mean backscatter values. This is evident in the tropical regions studied in [5] where, for example, the class of secondary forests, ranging in age from to years, exhibits a lot of variation. Actually this is also true in Flevoland to a certain extent. everal cover types, notably grasslands or fruit tree plantations, show large variation between fields. In [5], using Kolmogorov-mirnov (K-) tests of fit, it is shown that, for example, the Beta distribution is a better descriptor for the coherence magnitude, and the log-normal distribution for the backscatter level. In this paper this will be evaluated for a number of agricultural crop classes, grasslands and fruit tree plantations at the Flevoland test site, using AirAR (C-, L- and P-band polarimetric) data of the MAC Europe 99 campaign. A new reversible transform of the covariance matrix will be introduced in order to describe the full polarimetric target properties in an alternative way, allowing even simpler statistical descriptions. It will be shown that this transform yields versatile and robust classification approaches. Also the effect of azimuthally asymmetric backscatter behaviour on the classification results will be discussed. A comparison of results for the various classification methods will be given, using several (combinations) of frequency bands.. THEORY TRANFORM Fully polarimetric target properties can be described with nine independent numbers. For the covariance matrix C, () a Hermitian matrix, these properties are contained in the three real numbers on the diagonal and the six real and imaginary parts of the three complex numbers above the diagonal. Another way of presenting the full polarimetric information content, and perhaps better related to physical understanding, is using three intensity values, three polarisation phase difference values and three coherence magnitude values (with as the complex coherence) as π C. () It is also possible to describe the full polarimetric information content with nine intensities, for example as Im Re Im Re Im Re l hl h rr ll B, with 5 5 B () An elaboration is given in the Appendix. It is thus shown that the fully polarimetric radar target properties can also be collected by nine independent single-polarisation radar systems, measuring normalized radar cross-sections only. This

3 may not be very practical as a measurement approach, however is interesting from a theoretical point of view. It is also the theoretical basis for further improvement to be presented in this paper of the algorithm introduced initially by the author in [5]. Another application (or spin-off) is the possibility to make a polarimetric segmentation using nonpolarimetric segmentation software. An illustration is given in Fig.. Fig.. Four nonaplets showing a polarimetric segmentation application. In the first nonaplet the full polarimetric information content is shown (in normal reading sequence) as {,, ;,, ;,, }. In the second nonaplet the transformation to the system of Eq. is made: {,, rr, ll, ++ 5, 5, h+5, hl, +5l }. ince these are intensities only a segmentation can be made with conventional, i.e. non-polarimetric algorithms. The third nonaplet shows the result using ecognition. The fourth nonaplet is the result of applying the transform inversely. The fifth image in the fourth nonaplet, for example, shows a segmented HHVV phase difference image. The image shown is an L-band sub-set of 5 x 5 pixels dimension.. FLEVOLAND TET ITE AND DATA BAE For the AirAR image collected at July 99 a ground truth data set of agricultural fields is available. For the analysis presented in this paper only the -6 inc. angle range is used, for which 6 fields are available (Table ). At this date, in the middle of the growing season, the main crops are characterized as follows [6]. ugar beet fields have a cover of -6% and a height of -5 cm. Potato fields have a cover of 9-95% and a height of 5-6 cm. Wheat fields have a cover of 5-95% and a height of 5-95 cm. Volumetric soil moisture level varies between -%. Table. Cover types, codes and number of fields in the -6 inc. angle range. Barley BAR Oats OAT Beans BEA 6 Onions ONI Corn COR Peas PEA Flax FLA 6 Potato POT 6 Fruit trees FRU Rapeseed RAP Grassland GRA ugarbeet BT Lucerne LUZ Wheat WHE 5. CLAIFICATION APPROACH Certain fields may exhibit azimuthally asymmetric polarimetric behaviour [7]. uch behaviour is known from earlier experiments in Flevoland [, 9]. Examples of non-symmetric azimuthal behaviour are bending caused by prevailing winds (stem beans []), harvesting patterns [], lodging (by strong winds) [9], tillage patterns (prominent in potato [9]) or heliotropy (sunflowers). When not accounted for properly this may cause problems in the classification procedure. For example a potato field with ridge orientation 5º with respect to flight direction differs from an otherwise identical potato field with a 5º ridge orientation in its polarimetric signature. The extra information on ridge orientation may disturb the classification process. A solution may be to discard this information which is present in the four asymmetric elements of the covariance matrix. In case of azimuthal symmetry [9] the covariance matrix simplifies to

4 C r, () and. When interpreting remote sensing data it is sometimes useful to consider this and only this information, which is contained in 5 independent values of C. Also in the intensity representation introduced here it is possible to find several sets of 5 independent intensity values containing this and only this information. In all cases at least one composite (of the original used here) intensity is needed. One (non-redundant) possibility is Re Im [ ] [ ] B r l + 5r, with B r. (5) It is out of the scope of this paper to discuss all possibilities. However, there is one redundant system of particular interest which combines 7 intensity values, including two composite values (out of the original used here), namely {,,, +5, lr, + 5l + 5r, + 5r + 5l }. (6) Using Kolmogorov-mirnov (K-) tests of fit it is shown in [5] that for field averaged backscatter values the backscatter intensities are well described by the log-normal distribution, the (HHVV) phase difference by the circular Gaussian distribution and the (HHVV) coherence magnitude by the Beta distribution. After adding speckle, drawn from the complex Wishart distribution, this is still the case for the values of individual pixels belonging to a certain class. Note that for an individual homogeneous field the complex Wishart distribution, or its marginal distributions, are appropriate. However, for classification of a complex scene, featuring between field variations, it are the class distributions that are of primary importance. In analogy with [5] the K- tests were performed on the grassland, potato, sugar beet and wheat fields, and also on fruit tree plantations. In the latter case the incidence angle range was extended to º-7º, yielding a total of fields. The tests were done for all frequency bands, for the (HHVV) phase difference and coherence magnitude, and for all backscatter intensities used in equations, 5 and 6, including the two composite values. ome results for L-band and grassland are shown in Fig.. These are typical results. The significance Q of the null-hypothesis, stating that the observed field averaged values are drawn from the corresponding theoretical distribution is usually quite high. After adding speckle, deviations from the theoretical distribution may be detected by evaluation of the K- statistic D, which is defined as the maximum distance between the cumulative pdf s of the theoretical function and the observation (see Fig. ). At the db level of speckle (or -look data), for the intensity distributions tested, in only a few cases D exceeds a value of., and for the phase and coherence distributions tested, in only a few cases D exceeds a value of.6 (a similar result as reported in [5]). In this paper four classification models, named I+, 5I, 7I and 9I, will be evaluated: - (I+) The first model is identical to the one proposed in [5]. It uses a joint log-normal distribution for the HH-, HV- and VV-intensities and independent distributions for the phase difference (circular Gaussian distribution) and the coherence magnitude (Beta distribution). - (5I) The second model is based on Eq.5 and uses a joint log-normal distribution for the 5 intensities. - (7I) The third model is based on Eq.6 and uses a joint log-normal distribution for the 7 intensities. - (9I) The fourth model is based on Eq. and uses a joint log-normal distribution for the 9 intensities. Note that the first three models assume azimuthal symmetry or, alternatively, discard information related to azimuthal asymmetry. The fourth model includes all polarimetric information.

5 Fig.. Theoretical cumulative probability density distributions for grasslands for L-band HV- and VV-polarised backscatter, HHVV phase difference and coherence magnitude, respectively, compared with experimental observation (step-functions). The four graphs on the left are for field averaged values and have K- fit significance Q values of.,.59,.5 and., respectively. The four graphs on the right show the situation after adding speckle drawn from the -look complex Wishart distribution. These have K- test statistic D values of.9,.,.59 and., respectively. 5. REULT The classifiers are developed for the º-6º incidence angle range of the July 99 image (see also Table ). The effect of incidence angle dependence is ignored. To mitigate possible significant effects of incidence angle (cf. [5]) the backscatter parameter γ is used instead of ( γ / cos( θ i ) ). The evaluation of the results is also done on the º- 6º incidence angle range. everal pre-processing steps have been considered. The first is spatial aggregation of x pixels or x pixels. The second is segmentation. In this paper only results are shown using the latter pre-processing step, which gave slightly better results than for the x aggregates. The segmentation was done using the transform discussed in ection and ecognition (non-radar) segmentation software, as illustrated in Fig.. Fig. shows the AirAR image, crop type map, the classification result for a combination of C- and L-band using model 7I and an error map. A summary of the results for the different models and for different frequency band combinations is given in Table. The algorithm allows selection of a sub-set of the parameters used. ome results are given in Table. Table. Fully polarimetric classification results for several frequency band combinations for the º-6º incidence angle range of the July 99 AirAR image. Model Band Result Model Bands Result I+ C.5% I+ CL 95.% 5I C 9.% 5I CL 95.% 7I C 9.% 7I CL 96.% 9I C 6.% 9I CL 96.% I+ L.% I+ CLP 96.% 5I L 6.% 5I CLP 96.9% 7I L 6.% 7I CLP 97.% 9I L.7% 9I CLP 97.% Table. ingle- and multi-polarisation classification results for the C- and L-band for the º-6º incidence angle range of the July 99 AirAR image. Polarisation C-band L-band HH 9.9%.% VV.%.6% HH/HV 7.% 67.9% VV/HV 66.% 57.% VV/HH 6.% 6.% VV/HH/HV.9% 76.%

6 6º Fig.. (Top left) AirAR Total Power image (C-band blue; L- band green; P-band red) of Flevoland, July 99. (Top right) Crop type map. (Bottom left) Classification using C- and L- band full polarimetry model I7. (Bottom right) Error map: errors in red. The classification accuracy is 96.% for the area above the 6º inc. angle line (and 95.% for the whole image).

7 It is out of the scope of this paper to discuss all results in detail. A selection of the best results will be discussed next. The fully polarimetric overall classification result for the combination of C- and L-band using model 7I is very good (96.%). In absolute terms the main source of error is between wheat (acc. 9.%) and barley (acc. 96.%). There are only out of the cover types which have an accuracy less than 9%. These are fruit trees (.6%), which is confused with sugar beet; beans (6.7%), which is confused with grass, corn and sugar beet; onion (7.7%, which is confused with grass; and peas (.7%), which is confused with sugar beet. The overall classification result for C-band using model 7I is also very good (9.%). There are only four cover types with a classification result less than 5%. These are corn (67.%); fruit trees (.%), which confuses with potato and sugar beet; beans (7. %), which confuses mainly with sugar beet; and oats (79.6%), which confuses with corn. The overall classification result for L-band using model 9I is good (.7%). There are six cover types with a result below %. These are wheat (76.%), grass (7.7%), fruit trees (5.6%), beans (59.%), oats (7.%) and onion (6.%). The relatively bad result for beans in all cases (confusing with sugar beet) is suspected to be an error in the ground truth for one of the fields. 6. DICUION AND CONCLUION In general good overall classification results are obtained. C- band (9.%) performs a little better than the L-band (.7%) and both the C-L (96.%) and C-L-P (97.%) combinations have very little remaining error. It is interesting to compare the performance of the different models. In general system 7I is the best followed by 5I and I+. This result could be expected since all three models take the symmetric components, i.e. the same information, from the covariance matrix. However in I+ only the co-variances between the three intensity values are taken into account (and phase information is treated independently), while for the model 5I all co-variances (this time between 5 intensities) is taken into account. Model 7I improves upon 5I since the redundant elements seem to support a better (mathematical) conditioning of the same information. The 9I model can be better or worse than the other three models. The physical reason is not entirely clear, but is likely related to polarimetric asymmetry. uch effects may disturb or may improve classification when not accounted for properly. More research to further physical understanding of such effects is required. Also in comparison with the results cited in literature, as discussed in the introduction of this paper, the results are very good, notably for the C-band. A summary is given in Table. Of course an objective comparison is not possible since none of the methods is optimised or complete for mapping yet and different evaluation methods may have been used. Moreover, a different image is used here, earlier in the growing season. Nevertheless, it is the same area, which, in a statistical sense, has the same crop type and crop acreage composition. Table. Comparison with best results cited in literature (see ection ) for the same test site, the same radar system and for a single observation date. Ref. No. of classes PLC LC L C (excl. water) [] 9 9.5% [] 9.% [] 95.% 7.% 67.9% [] 9.%.6% 66.5% This paper 97.% 96.%.7% 9.% The good classification probably is the combined result of the use of realistic pdf s and exclusion of azimuthal asymmetry, though the latter point may require much more study before firm statements can be made. The transform to an intensity only system is very supportive as it allowed for the development of systems 5I and 7I which outperform the system I+, as introduced in [5]. In some cases system 9I is even better. Another asset of the transform is that it allows full polarimetric image segmentation using non-polarimetric segmentation software, such as multi-channel (intensity) radar segmentation algorithms. For multi-look data with a high number of looks (such as AirAR) even non-radar segmentation algorithms may suffice (as was done here). Further improvements may be expected when incidence angle effects are accounted for. It can be concluded that it seems possible to use RADARAT- and ALO PALAR for accurate crop type mapping and that a simple and robust technique as presented here may be sufficiently accurate. For operational crop monitoring systems time series will be used and results can be expected to improve further.

8 However, often these radar systems will not operate in the polarimetric mode and, like ENVIAT AAR, deliver single- or multi-polarisation data. Table gives an indication of the accuracy added by full polarimetric systems. In general in can be stated that the accuracy increases significantly when going from single polarisation (%-%), to double (57%-7%), triple (76%-%) and full polarisation (9%-9%). APPENDIX The full polarimetric information content can be described with nine intensity values in many ways. To evaluate this more systematically all combinations of the six antennae of the three common polarisation bases may be combined in a highly redundant system of 6 intensities. In the process of taking out redundancy some interesting features are revealed. In the backscatter direction, because of reciprocity, i.e. pq qp, only intensities are different. To reveal symmetry relations a schematic presentation is shown in Fig. where a division is made into nine sub-sets. Each subset thus contains intensities. When added the sum equals the Total Power (TP) for all nine sub-sets. h v l r h v l r h+ 5 h5 hl hr v+ 5 v5 vl vr l + 5r 5 5l 5r ll lr rr Fig.. chematic presentation of the set of intensities used and its division in nine sub-sets. There are 6 independent sub-set cases: ll + rr + lr TP TP h h5 + v+ 5 + v5 hl + hr + vl + vr TP + 5l + + 5r + 5l + 5r + + TP (7a) (7b) (7c) TP (7d) (7e) TP (7f) Further, summing the diagonal values, ll rr TP, (7g) which leads to the dependent relation lr + TP (7h) Using the 7 independent relations with TP the intensities can be reduced to 5 in many obvious ways. A further reduction to a complete set of nine intensities can be accomplished by some more complex relationships (or reduction rules), and/or through reduction of sets of linear equations. The latter may be done on any of the sets of 5 (as indicated above) or directly from the original set of. One solution out of the many was already given above (Eq.). Complete sets of nine intensities have certain interesting symmetry properties. However, it is out of the scope of this Appendix to treat these in more depth.

9 ACKNOWLEDGEMENT Martin Vissers of arvision BV is acknowledged for support in software development. REFERENCE [] Ferrazzoli, P., L. Guerriero, and G. chiavon, 999, Experimental and model invesigation on radar classification capability, IEEE Transactions on Geoscience and Remote ensing, Vol.7. No., pp [] Fukuda,., and H. Hirosawa, 999, A wavelet-based texture feature set applied to classification of multi-frequency polarimetric AR images, IEEE Transactions on Geoscience and Remote ensing, Vol.7. No.5, pp.-6. [] Chen, K.., W.P. Huang, D.H. Tsay, and F. Amar, 996, Classification of multi-frequency polarimetric AR imagery using a dynamic learning neural network, IEEE Transactions on Geoscience and Remote ensing, Vol.. No., pp.-. [] Lee, J.., M.R. Grunes, and E. Pottier,, Quantitative comparison of classification capability: Fully polarimetric versus dual and single-polarization AR, IEEE Transactions on Geoscience and Remote ensing, Vol.9. No., pp.-5. [5] Hoekman, D.H., and M.J. Quiñones,, Land cover type and biomass classification using AirAR data for evaluation of monitoring scenarios in the Colombian Amazon, IEEE Transactions on Geoscience and Remote ensing, Vol., pp [6] Vissers, M.A.M., and J.J. van der anden, 99, Ground truth collection for the JPL-AR and ER- campaign in Flevoland and The Veluwe (NL) 99, NRP- report 9-6, Delft, The Netherlands. [7] Nghiem,.V.,.H. Yueh, R. Kwok, and F.K. Li, 99, ymmetry properties in polarimetric remote sensing, Radio cience, Vol.7, pp [] Hoekman, D.H. and B.A.M. Bouman, 99, Interpretation of C-and X-band radar images over an agricultural area, the Flevoland test site in the Agriscatt-7 campaign, International Journal of Remote ensing, Vol., pp [9] Bouman, B.A.M. and D.H. Hoekman, 99, Multi-temporal, multi-frequency radar measurements of agricultural crops during the Agriscatt- campaign in the Netherlands, International Journal of Remote ensing, Vol., pp