Rajesh K Dhumal, YogeshD.Rajendra, Dr.K.V.Kale / International Journal of EngineeringResearch and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.758-761758 | P a g eClassification of Crops from remotely sensed Images: A reviewRajesh K Dhumal*, YogeshD.Rajendra**,Dr.K.V.Kale**,*Research Fellow, Department of Computer Science& IT, Dr.Babasaheb Ambedkar Marathwada University,Aurangabad (M.S.) India.**Research Fellow, Department of Computer Science& IT, Dr.Babasaheb Ambedkar Marathwada University,Aurangabad (M.S.) India.**Professor& Head, Department of Computer Science & IT, Dr.Babasaheb Ambedkar Marathwada University,Aurangabad (M.S.)Inda.ABSTRACTCrops identification from remotelysensed images is essential due to use of remotesensing images as an input for agricultural &economic planning by government & privateagencies, Available satellite sensors like AWIFS,LISS (IRS series), SPOT 5 and alsoLANDSAT,MODIS are good source ofmultispectral data with different spatialresolutions & Hyperion, Hy-Map, AVIRIS aregood source of hyper-spectral data. Expectedmethodology for this work is selection of satellitedata, use of suitable method for classification andchecking the accuracy of classification, from lastfour decades various Researcher has beenworked on this issues up to some extent but stillsome challenges are there like multiple cropsidentification, differentiation of crops of sametype this paper investigate the research workdone in this concern & critical analysis of thatwork has been presented here. Multispectral &hyper-spectral images contain spectralinformation about the crops good soft computing& analysis skills required to classify & identifythe class of interest from that datasets, variousresearchers has been worked with supervised &unsupervised classification along with hardclassifiers as well as soft computing techniqueslike fuzzy C mean, support vector machine &they found different results with differentdatasetsKeywords-Crops Classification, Hypers-pectral,Multispectral,Microwave, Remote sensing,MicrowaveI. INTRODUCTIONRemote sensing, particularly satellites offeran immense source of data for studying spatial andtemporal variability of the environmental parameters. Remote sensing has shown great promise inidentifying the crops grown in agricultural land. Theresultant information has been found to be useful inthe prediction of crop production and of land use. Itis playing a significant role for crop classification,crop health and yield assessment. Since the earlieststages of crop classification with digital remotesensing data, numerous approaches based onapplying supervised and unsupervised classificationtechniques have been used to map geographicdistributions of crops with optical data andCharacterize cropping practices Hyper spectral datacontains hypercube with no of bands which also actas a good source of information the smallestbandwidth of the hyper-spectral data tells fine detailabout the crops internal contents, not only theoptical but also Microwave remote sensing playing agood role for handling issues related to crops due toits distinct features.II. REMOTE SENSING FOR CROPCLASSIFICATIONClassification is the process where weconvert multilayer input image in to single layerthematic map, However, classifying remotely senseddata into a thematic map remains a challengebecause many factors, such as the complexity of thelandscape in a study area, selected remotely senseddata, and image-processing and classificationapproaches, may affect the success of aclassification.Broadly there are two approachesfor classification that are supervised classificationand Unsupervised classification as shown in Fig.1Unsupervised Classification is a clustering analysisin which pixel are grouped into certain categories interms of the similarity in their spectral values, in thisanalytical procedure all pixels in the input Data arecategorized into one of the groups specified by theanalyst beforehand. Prior to the classification theimage analyst does not have to know about scene orcovers to be produced During post processing eachspectral cluster get linked to meaningful Labelrelated to actual ground cover thus as shown in fig.1in Unsupervised Classification analyst just give thenumber of cluster & iteration then suitablealgorithms gives given no. of clusters for inputimage, Supervised Classification is much morecomplex than Unsupervised classification in thatAnalyst should aware about ground cover Process ofsupervised classification involve the selection ofappropriate band then Definition ofsignature fortraining samples these signature formsfoundation forthe subsequent classification care must be taken intheir selection. Selection of quality training samplesrequires knowledge of and understanding of the
Rajesh K Dhumal, YogeshD.Rajendra, Dr.K.V.Kale / International Journal of EngineeringResearch and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.758-761759 | P a g eproperties of the Different ground features in thesatellite imagery .Figure 1.General methodology for classificationof remotely sensed image(Supervised &Unsupervised).Visual interpretation during classificationis based on standard FCC (False Color Composite)generated using green, red and near-IR bandsassigned blue, green and red colors respectively.Each crop having their own distinct internalstructures, some crops may have similarities, anddue to distinctness of each crop they have differentspectral signatures it.is complicated to classify cropswith similar internal structure or similar reflectancebehavior in this case hyper-spectral imagery playsimportant role to find the minute difference betweenthese spectrally similar cropsIII. USE OF MULTISPECTRAL IMAGESMultispectral remote sensing systems useparallel sensor arrays that detect radiation in a smallnumber of broad wavelength bands Themultispectral airborne as well as satellite remotesensing technologies have been utilized as awidespread source for the purpose of remoteclassification of vegetation . Depending ongeographic area, crop diversity, field size, cropphonology, and soil condition, different band ratiosof multispectral data and Classifications schemeshave been applied  Nelis 1986 for example, useda maximum likelihood classification approach withLandsat data to map irrigated crop area in the U.S.High Plains. Further Priece in 1997 refined suchapproaches, using a multi-date Landsat ThematicMapper (TM) dataset in southwest Kansas to mapcrop distribution and USDA Conservation ReserveProgram (CRP) lands in an extensive irrigated area..Multi temporal data improve the accuracy ofclassification SujayDatta et.al. have used LISS 1Data for wheat crop classification by combiningJanuary & February two dates data & deriving therefirst two principle components they got 94 %classification accuracy . K.R.Manjunathet.al(1998) have been worked on IRS LISS II &LISS III for crops identification & conclude thatspatial resolution & spectral band selection areaffects on the classification results depending oncrops area. S.P.Vyaset.al. have been used multidate IRS LISS III data for multi crop identificationand crop area estimation in Utter Pradesh (India),They worked for mustard, potato & wheat cropsafter performing classification and comparing theresult with single date classification results theyconclude that with the use of multi date IRS LISSIII data it was possible to discriminate & map thevarious crops in the study area whereas single dateis still good for homogeneous area & reduce datavolumeI. P. B. C. Leiteat el(2008) has been worked onHidden Markov Model (HMM) based technique toclassify agricultural crops. They have used 12Landsat images for 5 crop types, indicated aremarkable superiority of the HMM-based methodover a mono temporal maximum likelihoodclassification approach. And found 94% accuracysimilarly they conclude that HMM approach alsoperformed well to recognize phonological stages ofcrops, Qiong An et al.(2009) have been useadaptive feature selection model In his work for ricecrop with MODIS data the extracted spectralcharacteristics are analyzed using statistical methodand dynamic changes of temporal series of indicesincluding NDVI, EVI, MSAVI and NDWI arestudied and by taking account of computationalcomplexity & time effectiveness of calculation theAdaptive Feature selection model(AFSM) is studied& he got 94 % accuracy which greater than generalclassification by 3%, VijayaMusande et. al. havebeen worked for Cotton crop discrimination usingfuzzy classifier approach in their work they haveused temporal AWIFS & LISS III data sets ofdifferent months as per life cycle of cotton crops andby generating indices like NDVI,TNDVI,SAVI &TVI they got improved vegetation signal.III. USE OF HYPER-SPECTRAL IMAGESHyper spectral remote sensing imagersacquire many, very narrow, contiguous spectralbands throughout the visible, near-infrared, mid-infrared, and thermal infrared portions of theelectromagnetic spectrum as shown in Fig.2 Hyperspectral data contains huge volumes so it iscomplicated to classify crops from it with sometraditional classification techniques, The availabilityof hyper spectral data has overcome the constraintsand limitations of low spectral and spatial resolution
Rajesh K Dhumal, YogeshD.Rajendra, Dr.K.V.Kale / International Journal of EngineeringResearch and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.758-761760 | P a g eimagery, and discreet spectral signature. Prasad S.Thenkabaiet.al(2002) have been worked onEvaluation of Narrowband and BroadbandVegetation Indices for Determining Optimal Hyperspectral Wave bands for Agricultural CropCharacterization in their works they have usedLandsat data with broad band narrow bands hyperspectral data with 430 bands in the visible & NIRFigure2.Conceptual Hyper spectral cube withcontinuous spectrum.portion of the spectrum, study has beenconducted for six crops namely barley, wheat,chickpea, lentil, vetch, and cumin by evaluating thevarious indices based on two bands they havechosen only 12 bands out of 430 which provideoptimal biophysical information about all 6crops some researcher proposed a procedure toreduce Dimensionality of hyper spectral data whilepreserving relevant information for posterior cropcover classification through Local Correlation&Sequential Floating Forward Selection algorithm(HFFS) .Xiangrong Zhanget.al(2009) havebeen worked for A new feature extraction methodbased on immune clonal selection (ICSA) and PCAfor classification of Hyper-spectral remote sensingimage& they conclude that given method givesbetter classification results.FaridMelganiandandLorenzoBruzzone have used support vector machinefor hyper-spectral image classification (2004) andthey conclude that SVMs Are effective than Radialbased function neural networks and the K-nnclassifier in terms of classification accuracy,computational time, and stability to parametersetting. SVMs seem more effective than otherpattern recognition approach based onthecombination of a feature extraction/selectionprocedure and a conventional classifier, SVMsexhibit low sensitivity to the Hughes phenomenon,resulting in an excellent approach to avoid theusually time-consuming phase required by anyfeature-reduction methodShwetanket al(2011)have used EO 1 Hyperion data for rice cropsclassification they have developed the spectrallibrary for rice crop and they performed supervisedclassification with preprocessing & withoutpreprocessing, Preprocessing involves radiometriccorrection, geometric correction and abnormal bandand pixels Detection and correction, they usedspectral angle mapper for classification & they got86.96% accuracy for without preprocessed data and89.33% accuracy for preprocessed data.IV. USE OF MICROWAVE REMOTE SENSINGIMAGESMicrowave remote sensing, usingmicrowave radiation with wavelengths from about 1Centimeters to a few tens of centimeters enableobservation in all type weather condition withoutany restriction by cloud or rain. that’s why it canpenetrate through cloud cover, haze, dust, and allbut the heaviest rainfall,This is one of theadvantages which is not possible with visible &infrared remote sensing, Microwave remote sensingprovides unique information for example sea windand wave direction which are derived fromfrequency characteristics, Doppler’s effect,Polarization, back scattering etc. that cannot beobserved by visible and infrared sensorsOpticalremote sensing is good source for cropsclassification in spite of that its limitation due toenvironmental interface like clouds which resultsinto scattering effect & we can’t get the fine detailsfrom that images in that case Microwave Remotesensing plays good role. Operating in microwaveregion of the electromagnetic spectrum improvessignal penetration within vegetation & soil targetsthe longer wavelength of the RADAR system arenot affected by cloud cover or haze, RADAR systemtransmits microwave signal at specific wavelengthaccording to their design specification .Jesus SoriaRuiz et.al.(2009) have been worked on cornmonitoring & crop yield using microwave & opticalremote sensing & conclude that integration of SAR& optical data can improve the classificationaccuracy, Jiali Shang et.al. (2010) have beenworked onMulti-temporal RADARSAT-2 andTerraSAR-X SAR data for crop mapping in Canada& they found that When multi-frequency SAR (X-and C-band) are combined, classification accuraciesabove 85% are achieved prior to the end of seasoncrops can be identified with accuracies between86% (western Canada) and 91.4% (easternCanada).V. CONCLUSIONRemote sensing images are act as goodsource for decision making related to cropsmonitoring & mapping in optical remote sensingmultispectral images gives much detail for overallvegetation mapping in large area whereas its havinglimitation due broad wavelength & spatial resolutionwe can’t differentiate crops of similar type in thatcase hyperspectral images doing well, Selection ofspectral bands in hyperspectral images is also quite
Rajesh K Dhumal, YogeshD.Rajendra, Dr.K.V.Kale / International Journal of EngineeringResearch and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.758-761761 | P a g echallenging task whereas applying various hard aswell as soft classifiers gives good classificationresults, some limitation of optical remote sensingcan be overcome by fusing optical remote sensingimages with microwave remote sensing images.ACKNOWLEDGEMENTSWe thanks to Ramanujan Geospatial Chairestablished at Department of Computer Science &IT, Dr. Babasaheb Ambedkar MarathwadaUniversity Aurangabad (M.S), and India. for givingus opportunity to work in this area of researchREFERENCES Perumal, K., and R. Bhaskaran."Supervised classification performance ofmultispectral images." arXiv preprintarXiv:1002.4046 (2010).] Lu, D., and Q. Weng. "A survey of imageclassification methods and techniques forimproving classificationperformance." International Journal ofRemote Sensing 28.5 (2007): 823-870. Landgrebe, David. "Information extractionprinciples and methods for multispectraland hyperspectral image data." Informationprocessing for remote sensing 82 (1999):3-38. Nellis, M. Duane, Kevin P. Price, andDonald Rundquist. "Remote sensing ofcropland agriculture." The SAGEHandbook of Remote Sensing 1 (2009):368-380. Dutta, Sujay, et al. "Wheat cropclassification using multidate IRS LISS-Idata."Journal of the Indian Society ofRemote Sensing 26.1 (1998): 7-14. Manjunath, K. R., N. Kundu, and S.Panigrahy. "Evaluation of spectral bandsand spatial resolution of LISS II and LISSIII sensors on-board IRS satellites for cropidentification." Journal of the IndianSociety of Remote Sensing 26.4 (1998):197-208. Vyas, S. P., M. P. Oza, and V. K. Dadhwal."Multi-crop separability study of Rabicrops using multi-temporal satellitedata." Journal of the Indian Society ofRemote Sensing 33.1 (2005): 75-79. Leite, P., et al. "Crop type recognitionbased on Hidden Markov Models of plantphenology." Computer Graphics and ImageProcessing, 2008.SIBGRAPI08.XXIBrazilian Symposium on.IEEE, 2008. An, Qiong, et al. "Research on featureselection method oriented to cropidentification using remote sensing imageclassification." Fuzzy Systems andKnowledge Discovery, 2009.FSKD09.SixthInternational Conference on.Vol. 5.IEEE,2009. Musande, Vijaya, Anil Kumar, andKarbhari Kale."Cotton CropDiscrimination Using Fuzzy ClassificationApproach." Journal of the Indian Society ofRemote Sensing (2012): 1-9. Thenkabail, Prasad S., Ronald B. Smith,and Eddy De Pauw. "Evaluation ofnarrowband and broadband vegetationindices for determining optimalhyperspectral wavebands for agriculturalcrop characterization." PE & RS-Photogrammetric Engineering & RemoteSensing 68.6 (2002): 607-621. Gomez-Chova, L., et al. "Feature selectionof hyperspectral data through localcorrelation and SFFS for cropclassification." Geoscience and RemoteSensing Symposium,2003.IGARSS03.Proceedings.2003 IEEEInternational.Vol. 1.IEEE, 2003. Xiangrong Zhang, Runxin Li and LichengJiao “Feature Extraction Combining PCAand Immune Clonal Selection forHyperspectral Remote Sensing ImageClassification”, International Conferenceon Artificial Intelligence andComputational Intelligence,2009. Melgani, Farid, and Lorenzo Bruzzone."Classification of hyperspectral remotesensing images with support vectormachines." Geoscience and RemoteSensing, IEEE Transactions on 42.8(2004): 1778-1790. Shwetank, S., Kamal Jain, and KaramjitBhatia. "Development of Digital SpectralLibrary and Supervised Classification ofRice Crop Varieties Using HyperspectralImage Processing." Asian Journal ofGeoinformatics 11.3 (2012). Ruiz, Jesus Soria, Yolanda FernandezOrdonez, and Heather McNairn. "CornMonitoring and Crop Yield Using Opticaland Microwave Remote Sensing." Shang, Jiali, et al. "Contribution ofTerraSAR-X Data to In-Season CropMapping in Canada." Proceedings ofTerraSAR-X Science Team Meetinghttp://www.nature.com/nphoton/journal/v3/n11/fig_tab/nphoton.2009.205_F3.html.(Image 2 source). Gao, Jay. Digital analysis of remotelysensed imagery.McGraw-Hill Professional,2008. Shippert, Peg. "Introduction tohyperspectral image analysis." OnlineJournal of Space Communication 3 (2003).