IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 41, NO. 7, JULY 2003 1611 Crop Classification Using Multiconfiguration C-Band SAR Data Fabio Del Frate, Member, IEEE, Giovanni Schiavon, Domenico Solimini, Member, IEEE, Maurice Borgeaud, Senior Member, IEEE, Dirk H. Hoekman, Member, IEEE, and Martin A. M. Vissers Abstract—This paper reports on an investigation aimed at evaluating the performance of a neural-network based crop classi- fication technique, which makes use of backscattering coefficients measured in different C-band synthetic aperture radar (SAR) configurations (multipolarization/multitemporal). To this end, C-band AirSAR and European Remote Sensing Satellite (ERS) data collected on the Flevoland site, extracted from the European RAdar-Optical Research Assemblage (ERA-ORA) library, have been used. The results obtained in classifying seven types of crops are discussed on the basis of the computed confusion matrices. The effect of increasing the number of polarizations and/or measurements dates are discussed and a scheme of interyear dynamic classification of five crop types is considered. Index Terms—Crop classification, neural networks, synthetic aperture radar (SAR). I. INTRODUCTION T HE POTENTIAL of synthetic aperture radar (SAR) in dis- criminating among different agricultural crop species has been demonstrated in several studies [1]–[3]. The accuracy of classification depends on the sensitivity of the used backscat- tering coefficients to the differences of the biomorphological structures of the plants, hence to the different interaction be- havior between the electromagnetic wave and the structure of the canopy [4]. It has been experienced that measurements taken by a SAR system in a single configuration, that is one image at given fre- quency, polarization and incidence angle, are often inadequate to attain the required accuracy of classification. Given the de- pendence of the scattering mechanisms in vegetation canopies on frequency, polarization and incidence angle, improvements are expected by multifrequency and/or multipolarization and/or multiangle measurements [5]–[8]. Alternatively, multitemporal single-frequency, single-polarization data collected by repeated Manuscript received April 2, 2002; revised March 27, 2003. This work has been partially supported by Agenzia Spaziale Italiana (ASI). The data have been made available through the ERA-ORA Concerted Action, funded by the EC under Contract ENV4-CT97-0465. F. Del Frate, G. Schiavon, and D. Solimini are with the Dipartimento di Informatica, Sistemi e Produzione (DISP), Università Tor Vergata, I-00133 Roma, Italy (e-mail: delfrate@disp.uniroma2.it; schiavon@disp.uniroma2.it; solimini@disp.uniroma2.it). M. Borgeaud is with the Swiss Space Office, CH-3003 Bern, Switzerland (e-mail: maurice.borgeaud@sso.admin.ch). D. H. Hoekman and M. A. M. Vissers are with the Department of En- vironmental Sciences, Wageningen University, 6709 PA Wageningen, The Netherlands (e-mail: dirk.hoekman@users.whh.wau.nl; martin.vissers@ users.whh.wau.nl). Digital Object Identifier 10.1109/TGRS.2003.813530 overpasses can improve the accuracy, since they are affected by the peculiar variations induced in backscattering by the growth cycle of a given plant [9]–[11]. To be successful, suitable classification algorithms should be used, which are capable of exploiting the information embedded in multipolarization and multitemporal SAR measurements. A variety of classification schemes have been proposed and used, some recent examples of which can be found in [12]–[14]. Due to several interesting and peculiar features, neural net- work algorithms (NNAs) have also been considered for classi- fication purposes [15]–[17]. With respect to statistical methods, neural networks use an essentially different approach, so that they do not rely on probabilistic assumptions neither need par- ticular requirements about normality in datasets. This paper reports on an investigation aimed at a systematic evaluation of the information content, hence of the classifica- tion potential, of different consistent sets of C-band backscat- tering coefficients of agricultural fields. Multipolarization data consist of the set of measurements collected over Flevoland, The Netherlands, by the National Aeronautics and Space Ad- ministration Jet Propulsion Laboratory (NASA/JPL) AirSAR system during the 1991 MAC-Europe campaign, while mul- titemporal data over the same site were acquired by the Eu- ropean Remote Sensing Satellite 1 (ERS-1) SAR in the years 1993, 1994, and 1995. The data used in this study have been extracted from the European Radar-Optical Research Assem- blage (ERA-ORA) Library, assembled through a concerted ac- tion funded by the European Commission within the Research and Technology Development Programme on Environment and Climate (Fourth Framework Programme) in the field of space techniques applied to environmental monitoring and research [18]. The classification algorithm, consisting of a multilayer neural network with feedforward configuration, has been fed by sets of data of varying completeness. The corresponding varia- tion of classification accuracies of selected crop species, as ex- pressed by the confusion matrices, is discussed and related to the type of input measurements. The results obtained by the neural net are compared with those of a maximum likelihood algorithm [8]. A dynamic classification scheme, aimed at discriminating the crops during their development phase is also presented and examined. Our classification exercise makes use of data taken at high incidence angles. This choice reduces the possibly detrimental effects of the underlying soil, but, from a practical point of view, it would limit to the far range the portion of an airborne SAR image over which classification is expected to be performed ef- fectively. Extending the area requires the adoption of suitable 0196-2892/03$17.00 © 2003 IEEE