rice monitoring in spain by means of time series of terrasar-x dual ...

RICE MONITORING IN SPAIN BY MEANS OF TIME SERIES OF TERRASAR-XDUAL-POL IMAGESJuan M. Lopez-Sanchez 1 , J. David Ballester-Berman 1 , and Irena Hajnsek 21 Signals, Systems and Telecommunications Group, EPS, University of Alicante, P.O. Box 99, E-03080 Alicante, Spain,Phone: +34 965909597, E-mail: juanma.lopez@ua.es, davidb@ua.es2 German Aerospace Center (DLR), Microwaves and Radar Institute P.O. Box 1116, 82234 Wessling, Germany,Phone: +49-(0)8153-28-2363 E-mail: irena.hajnsek@dlr.deABSTRACTThree time series of dual-pol TerraSAR-X images havebeen acquired during the whole cultivation period for arice site in Spain. Among different observations, thebackscattering at HH and VV channels, and the HH/VVratio at 30 degrees incidence has shown a temporal signaturewith a clear correlation with the development ofthe plants during the vegetative and reproductive phenologicalphases. A physical interpretation has been providedfor this response. Results provided by the dual-polentropy-alpha decomposition and by the phase differencebetween copolar channels have been also introduced andpreliminarily discussed. In addition, these images havedemonstrated to be useful for discriminating rice fieldsfrom other crop types and also for detecting areas withirrigation problems inside a rice field.1. INTRODUCTIONThe launch of the TerraSAR-X satellite sensor opened anew spectrum of potential applications in many remotesensing fields. Regarding agriculture monitoring, themain novelties provided by this satellite are the following:• Frequency band: For the first time it is possible toanalyze time series of X-band images acquired bya satellite in continuous operation. Compared toC or L-band, this frequency may be better adaptedto scenes with short vegetation (such as agriculturalcrops) because the backscattering response from theplants and the ground is balanced.• Dual polarization preserving phase information: Polarimetricchannels acquired in dual-pol mode maintaintheir relative phase, so they can be combinedcoherently. Moreover, the transmitting polarizationcan be different for both channels, so imageswith HH and VV simultaneously acquired are available,in addition to the more common VH/VV andHV/HH combinations.• Revisit time: The cycle of the satellite is 11 days,much shorter than previous sensors, so it is possibleto build time series of observables with lessdramatic changes between successive acquisitions.This feature is specially important for agriculture,since crops change very rapidly and the whole phenologicalcycle covers from 1.5 to 4 months, dependingon the crop type.• Spatial resolution: A high spatial resolution (2 x 6 min stripmap dual-pol mode) is mandatory to observeagricultural fields with small extensions and alsoto provide sufficient number of samples for multilooking.In addition, cultivation problems (such asplagues) affecting a part of a field can be detected.The specific monitorization of rice fields had been previouslyapproached with the available satellite sensors atC-band (ERS, Envisat-ASAR and Radarsat-1), and somesuccessful results have been reported in the literature(see, for instance, [1, 2, 3]) despite the limitations dueto sensors: long revisit time, single polarization, coarsespatial resolution, etc. There exist also some examples ofX-band experiments over rice fields, but they were conductedwith an airborne experimental sensor [4] and aground-based scatterometer [5].In this work, we present the first results of a campaigndesigned to evaluate the potential of the new TerraSAR-X sensor for the monitorization of rice fields. To this end,images acquired during the whole crop cycle at differentpolarization channels and from different incidence angleshave been processed.The first objective of this research is to identify observableswith clear signatures as a function of rice phenology,so the emphasis will be on the analysis of time seriesof most of the possible parameters that can be computedfrom the images (note that not all of them will be representedin this paper due to space constraints). Thoseparameters which show a signature with the phenologicalevolution of the rice crop will be interpreted in termsof the scene characteristics and the scattering processesinvolved. Finally, practical applications of this monitorizationwill be discussed.

MAY JUN JUL AUG SEP OCTNOVVH/VV22ºHH/VV 30ºXVH/VV40ºXX X X XSowingHarvestFigure 1. Schedule of acquisitions organized as 3 seriesof images. Black circles denote acquired images, andcrosses denote images which were ordered but not acquireddue to maintenance operations and collision withorders from another users2. SAR DATA AND GROUND TRUTH CAM-PAIGNThe test site consists of an area of 30 km x 30 km in themouth of the Guadalquivir river, SW of Spain, where riceis cultivated annually from May to October.During the 2008 campaign, the local association of ricefarmers (Federacion de Arroceros de Sevilla) has collectedseveral ground truth parameters on a weekly basis.For this research project, 8 specific parcels, spreadover the whole area, were selected for ground-truth monitorization.The area of each parcel is around 8–14 ha.Phenological stage and vegetation height on these parcelshave been annotated during the whole period. In addition,specific aspects for some of them have been registered,such as irrigation conditions, water salinity and presenceof plagues. Finally, there is also climate information providedby the Spanish Agency of Meteorology (AEMET),including daily files of temperature, precipitations, humidityand wind.In order to maximize the investigation of the capabilitiesof the TerraSAR-X sensor, the acquisition of imageswas organized as a set of three parallel time series. Eachtime series, with an 11-days revisit period, correspondsto dual-pol stripmap images obtained with a different incidenceangle, namely 22, 30 and 40 degrees. Note thatthe operational range of incidence angles of TerraSAR-Xcomprises 20 to 45 degrees. In addition, different polarizationcombinations were chosen for the time series:VH/VV for two of them (22 and 40 degrees) and HH/VVfor the 30 degrees case. The schedule of the acquisitionsis illustrated in Fig. 1, where the dates employed for sowingand harvest are also indicated. Note that sowing isnot simultaneous in all parcels of the site. In addition,the time span dedicated to harvest is about one month becausesome fields exhibit different development periodsand because there were some rain events in the last dates,which complicated the harvest and post-harvest practices.The temporal gaps between acquisitions are only 2, 4 and5 days, thus providing a very frequent coverage of the testsite.As a consequence of the dual-pol operation, the swath ofeach image is half the swath of single-pol case, so thefinal area covered by each image is about 15 km (ground-Figure 2. Temporal evolution of the backscattering coefficientsfor HH and VV for a single rice field (parcel #34)at 30 ◦ incidencerange) x 40 km (azimuth). Therefore, the whole rice areais not fully covered by the images, but the center region iscovered by all images (note that two series are descendingand one is ascending).3. TEMPORAL EVOLUTION OF POLSAR OB-SERVABLESWith the objective of facilitating the analysis of the observations,we start this section by showing the results for asingle parcel. The temporal evolution of the backscatteringcoefficients (σ 0 ) at HH and VV channels at 30 ◦ incidenceis plotted in Fig. 2. The vertical line at DoY 156denotes the date of sowing.Backscattering from the field is very low just after sowingat both polarizations because the ground was completelyflooded and, consequently, all the incident radarsignal is scattered in the specular direction (see blackareas in Fig. 3). From that moment, backscattering increasesas the vegetation grows, and the HH channel providesa higher response than VV, reaching a maximumon July 13. Then, from this date, the radar signal decreasesfor both channels, but the decrease rate is alsomore pronounced for vertical polarization. In fact, in theVV images acquired on August 26 and September 6 theaspect of the field is the same as the first image after sowing(see Fig. 3), i.e. extremely low backscatter, despitethe fact that the vegetation is fully developed.From September 6 to September 17, both backscatteringchannels increase slightly. The images acquired onSeptember 28 exhibit a very high backscattering in thewhole area. After consulting the weather archive, it wasnoticed that there was a strong rain the day before (55mm/m 2 ), and also some rain during the acquisition day.As a result of the rain, the water content in the vegetationvolume increases, and this is the reason of a higherbackscattering. More details about this interpretation areincluded below.

10-Jun-2008 DoY = 16113-Jul-2008 DoY = 194Figure 4. Daily precipitations registered at a meteorologicalstation (Isla Mayor) located in the center of the testsite26-Aug-2008 DoY = 238HHVVFigure 3. Images of σ 0 at HH (left) and VV (right) channelsat 30 ◦ acquired at three dates (from top to bottom)Unfortunately, during most of image acquisitions afterOctober 1, there were rain events and, sometimes, theywere accompanied by moderate winds (see Fig. 4). Consequently,it is very difficult to interpret this last stage ofthe cultivation campaign.In addition to the observation of HH and VV, it is interestingto study the ratio between both channel powers(HH/VV) as a function of time. The HH/VV ratio (seeFig. 5) increases in the first stages from 0 to more than 10dB. Then it decreases down to 2 dB, and it presents someoscillations in the last dates.When dealing with backscattering coefficients, the productsof TerraSAR-X can be calibrated in two differentways. The conventional calibration [6] allows the translationof DN (for detected products) or complex values(for single-look complex products, as in this case) to β 0and σ 0 , and can be performed on a pixel-by-pixel basis.The second possible calibration, also named radiometricenhanced (RE) [7], includes also a statistical subtractionof a noise estimation annotated in the products. The REcalibration is necessary when working with low backscatteringvalues, specially those close or lower to the noiseFigure 5. Temporal evolution of the HH/VV ratio for asingle rice field (parcel #34) at 30 ◦ incidencelevel. The drawback of the RE calibration is that it canbe only performed on multi-looked power images, so thephase information is not available anymore in the calibratedresults and spatial resolution is slightly degraded.The backscattering coefficients after the RE calibrationare shown in Fig. 6 for the same case of Fig. 2. We canobserve that differences are only evident for σ 0 valuesbelow -15 dB. This was expected because the NESZ isaround -19 dB for these images.In the rest of the paper we have used only the conventionalcalibration because we are interested in exploitingalso the phase information provided by the sensor. Ofcourse, in cases where low σ 0 values are requested fora particular application and phase is not necessary, theRE calibrated data should be used instead. Note also thatthe differences between the HH/VV ratio of the RE calibrateddata and the ratio provided by the conventionalcalibration are only evident at the points with very lowVV backscattering.After observing the response at VV and HH for one parcel,we show now in Figs. 7 and 8 the evolutions ofboth backscattering coefficients and their ratio for all the

Figure 6. Temporal evolution of the radiometrically enhanced(RE) backscattering coefficients for HH and VVfor the same rice field (parcel #34) at 30 ◦ incidenceFigure 8. Temporal evolution of the HH/VV ratio for allthe monitored parcels at 30 ◦ incidencetion height increases from 0 to about 70-90 cm. Then,the appearance of the panicle defines the transition to thereproductive stage. In the reproductive stage there alsoappears a cluster of curved leaves in the top of the plants,as well as flowering elements. Moreover, the stems becomedrier and tend to present a cane structure. Finally,in the mature phase the grains appear and develop. Thewater content of the grains changes up and down until itis about 18-22 percent before harvesting. For parcel #34,the vegetative phase finished about day 225 of the year(mid August), and the mature phase started around day270 (end of September).The influence of phenology in the radar responses can beinterpreted as explained in the following:Figure 7. Temporal evolution of the HH and VV backscatteringcoefficients for all the monitored parcels at 30 ◦ incidenceparcels with available ground truth. Since the sowing datewas different for the parcels, the time series are not plottedversus the date (DoY), but shifting them to a commontime origin at sowing date.It is important to stress that the responses of all parcelsbut one (plotted in red) are very coincident, both for theindividual σ 0 and the HH/VV ratio. Therefore, we canconclude that these temporal signatures are associatedwith this rice crop type and, in principle, can be used fordescribing its phenological evolution. The peculiaritiesof the outlier parcel (in red) will be explained in the nextsection.The interpretation of the overall temporal evolution canbe carried out by taking into account the phenologicalevolution of the plants. In general, the cultivation periodis divided into 3 main phenological phases: vegetative,reproductive and mature [8]. During the vegetative stagethe plant is composed of up to 10 stems (the main oneand new ones appearing during this period), and someleaves at the top in the late part of this phase. Vegeta-Vegetative phase The HH/VV ratio increases with timefrom 0 after sowing to around 8-12 dB, just at theend of the vegetative phase. Three considerationsdefine this response:1. Radar backscattering at X-band and 30 degreesincidence is dominated by the double-bounceinteraction between the flooded soil and thequasi-vertical stems. For this scattering mechanism,if extinction is ignored, it is known that:(a) When the stems are not very thin, totalbackscattering at HH is higher than at VVdue to the presence of the two bounces,with their corresponding Brewster angles,which attenuate the VV echo more thanthe HH one.(b) For very thin stems (like in this scene) thefinal values of HH and VV depend importantlyon the incidence angle and other parameters(such as the dielectric constantsof stems and water). This is a consequenceof the weak scattering at horizontalpolarization for such thin stems.(c) In general, when the double-bounce dominates,the value of the HH/VV ratio is

almost independent from the vegetationheight.2. For a scene composed by almost vertical verythin cylinders, extinction coefficient at verticalpolarization is larger than at horizontal polarization,i.e. there exists differential extinction.Again, the values of the extinction coefficientsdo not depend on vegetation height.3. As vegetation height increases (plants grow),the total attenuation is stronger for verticalthan for horizontal polarization, due to longerpropagation paths inside the vegetation volumeand the aforementioned differential extinction,and this is the mechanism driving the observedresponse. Note that attenuation at vertical polarizationis so strong that the backscattering atthe end of the vegetative stage (with already 90cm high plants) is around the noise level, likeat the time of flooded fields.Reproductive phase In this stage, the backscatteringvalues at both polarizations approach each other, asa result of the following aspects:1. The stems become drier than before (likecanes), so the reduced water content inducesless vertical extinction and, as a result, asmaller differential extinction. This entailsa progressive increase in the received VVbackscattering.2. The top part of the plants contain more elements(panicle, leaves, etc.) than in the vegetativephase. These elements are not vertical, butrandomly oriented and bent, so there is morescattering at this layer, and it is not as polarizationdriven as before.Maturation phase In principle, independence from polarizationis expected at this stage before harvest,since the plants become mostly a random volume.However, this has not been confirmed during thiscampaign due to the presence of strong rain andwind at this stage (see Fig. 4), which have alteredimportantly their radar signatures.As already mentioned, we are also interested in analyzingthe phase information provided by this new sensor.Fig. 9 presents the evolution of the phase difference betweenboth copolar channels. This parameter (usuallynamed as polarization phase difference, PPD, in the literature[9]), is computed as the phase of the complex crosscorrelation between both channels. In this case, it is quiteevident that for all fields there exists a jump in the phaseat 30 to 40 days after sowing, when the plants are visibleenough above the water level. This would be used as anindicator of the emergence time for the plants. In addition,for all fields the phase difference exhibits a plateauuntil day 90-110, i.e. until the end of the vegetative phaseapproximately.Maps of the PPD for a part of the rice test site are alsoshown in Fig. 10, corresponding to May 30 (about sowingFigure 9. Temporal evolution of the phase difference betweenVV and HH for all the monitored parcels at 30 ◦incidenceFigure 10. Images of the PPD from a part of the test site.Dates: Top - May 30. Bottom - August 26date) and August 26. The presence of the developed ricefields is quite evident in the second image, whereas thereis no PPD differences between fields in the first one.Thanks to the availability of phase information betweenthe copolar channels, we can also generate images for thefirst two elements of the Pauli basis, HH+VV and HH-VV, which are usually identified to direct and dihedraltypescattering. The result for the same parcel as beforeis shown in Fig. 11. In this case, however, the low levelof VV backscattering when compared to HH makes bothPauli scattering mechanisms very similar, and hence bothare approximately equal to HH.The angular diversity provided for the VV channel bythe three series of images has been also exploited here toshow the influence of the incidence angle on this channel.As observed in Fig. 12, at the very early stages after sowing,backscattering is higher for oblique incidences sincethe presence of very short plants can be observed betterat oblique angles. However, at the end of the vegetativephase, the order is the opposite, due to a stronger totalattenuation at more oblique incidences because the pathsof the waves inside the vegetation volume are longer foroblique angles.In addition to the previous results, it is possible to applyto these dual-pol images a polarimetric decomposition

and a decrease during the reproductive phase. A physicalinterpretation of these signatures, based on scatteringand extinction mechanisms present in the scene, has beenprovided. Other observations, like the phase differencebetween channels, the angular variation of backscattering,and the dual-pol entropy-alpha decomposition havebeen also tested. In addition, a quick identification ofthe cultivated rice fields and the detection of areas withcultivation problems have been proposed as possible applicationsof this remote sensing tool.The next steps in this project will cover different objectives.First, an electromagnetic model of the scene will bedeveloped in order to better understand the radar responseand to enable the construction of future algorithms for theretrieval of biophysical parameters of interest. To date, apreliminary model for the first part of the vegetative phasebased on Monte Carlo simulations [11], with the plantsmodelled only by the stems, have been implemented andtested. The addition of a layer of leaves, a better descriptionof the plants morphology by changing it as a functionof phenology stage, and the incorporation of multiplescattering to simulate clustering effects are currently beingincorporated to the model. Second, the automation ofthe image processing required for the proposed applications(e.g. rice fields identification, problems detection,etc.) is required. Third, the correlation of the observableswith the precise phenological phase and with theavailable biophysical parameters, provided by the groundtruth campaign, has to be conducted. The new observationspresented in this paper (phase differences betweencopolar channels and dual-pol entropy-alpha decompositionoutputs) should be also interpreted in terms of thescattering mechanisms present in the scene.Despite the fast growth of this type of crop, SAR interferometryobservables (coherence and phase) obtained bycombining images separated 11 days may provide alsohelpful information, so this will be also tested.Finally, a plan for the 2009 campaign has to be defined inthe framework of the same project. Among the decisionsto be taken, the selection of polarization channels and thetrade-off between spatial and temporal coverage will bestudied by attending the requests of the rice farmers.ACKNOWLEDGMENTSAll SAR images have been provided by DLR in theframework of project LAN0021 of the pre-launch AO ofTerraSAR-X.This work has been supported by the Spanish Ministryof Education and Science (MEC) under projectsTEC2005-06863-C02-02, TEC2008-06764-C02-02, andHA2007-075, by the Generalitat Valenciana underproject GVPRE/2008/344, and by the University of Alicanteunder project GRE08J01.The authors would like to thank the support of ManuelCano, Fernando Carrascal and Santiago Aparicio, allfrom the Federacion de Arroceros de Sevilla, for providingthe ground-truth data.REFERENCES[1] T. Kurosu, M. Fujita, and K. Chiba. Monitoringof rice crop growth from space using the ERS-1 C-band SAR. IEEE Trans. Geosci. Remote Sensing,33(4):1092–1096, July 1995.[2] T. Le Toan, F. Ribbes, L.-F. Wang, N. Floury, K.-H.Ding, J. A. Kong, M. Fujita, and T. Kurosu. Ricecrop mapping and monitoring using ERS-1 databased on experiment and modeling results. IEEETrans. Geosci. Remote Sensing, 35(1):41–56, January1997.[3] J.-Y. Koay et al. Paddy fields as electrically densemedia: Theoretical modeling and measurementcomparisons. IEEE Trans. Geosci. Remote Sensing,45(9):2837–2849, September 2007.[4] T. Le Toan, H. Laur, E. Mougin, and A. Lopes.Multitemporal and dual-polarization observationsof agricultural vegetation covers by X-band SARimages. IEEE Trans. Geosci. Remote Sensing,27(6):709–718, November 1989.[5] Y. Inoue et al. Season-long daily measurementsof multifrequency (Ka, Ku, X, C, and L) and fullpolarizationbackscatter signatures over paddy ricefield and their relationship with biological variables.Remote Sensing of Environment, 81:194–204, 2002.[6] T. Fritz et al. TerraSAR-X Ground Segment. Level1b Product Format Specification. Annex B. DLR.Doc. TX-GS-DD-3307, 2007.[7] T. Fritz. Tips & tricks for spectral, radiometric andgeometric interpretation of TerraSAR-X products.Part I. In Proceedings of the 3rd TerraSAR-X ScienceTeam Meeting, Wessling, Germany, November2008.[8] U. Meier, editor. Growth Stages ofMono- and Dicotyledonous Plants.BBCH Monograph. 2nd edition, 2001.http://www.bba.de/veroeff/bbch/bbcheng.pdf.[9] F. T. Ulaby, D. N. Held, M. C. Dobson, K. C. Mc-Donald, and T. B. A. Senior. Relating polarizationphase difference of SAR signals to scene properties.IEEE Trans. Geosci. Remote Sensing, 25(1):83–92,January 1987.[10] S. R. Cloude. The dual polarization entropy/alphadecomposition: a PALSAR case study. In Proceedingsof the 3rd PolInSAR Workshop, Frascati, Italy,January 2007.[11] L. Tsang, K. Ding, G. Zhang, and J. A. Kong.Backscattering enhancement and clustering effectsof randomly distributed dielectric cilinders overlyinga dielectric half space based on Monte-Carlosimulations. 43(5):488–499, May 1995.