OLSAR to classify forest types, and concluded that theuse of HH, HV, and VV polarization information improveddifferentiation between forest types without leaves. Rah-man and Sumantyo (2010) noted that vegetation informa-tion needed to differentiate between forest and non-foresttypes can be easily identified from Synthetic ApertureRadar (SAR) images. However, parameters extracted fromscattering matrix, covariance matrix, and correlation matrixdo not provide sufficient information for accurate POLSARimage classification in certain scenarios such as in complexforest areas where different scattering media exhibit asimilar POLSAR response for unavoidable reasons. Pre-vious studies indicated that polarization SAR textureinformation helps in improving classification results (Bor-ghys et al. 2006; Masjedi et al. 2016). POLSAR data arealso extensively used for terrain classification applyingSAR features from various target decompositions andcertain textural features (Uhlmann and Kiranyaz 2014). Adual-season POLSAR achieved the highest accuracies,suggesting that seasonality is critical to obtaining highaccuracy in wetland cover classification, irrespective of thetype of SAR image used (Furtado et al. 2016). TwoRadarsat-2 images acquired in leaf-on and leaf-off seasonswere selected for forest classification and found to beeffective (Maghsoudi et al. 2013).Recently, the development of remote sensing technol-ogy and application of earth observation satellite sensortechnologies has led to remote sensing offering multipleplatforms, multiple sensors, and multispectral characteris-tics and providing better spatial and temporal resolution.The fusion of different remote sensing technologiesenables classification of forest types with improved accu-racy. Kasapoglu et al. (2012) used fusion data from ALOSPALSAR and TM to classify forest types and documentedan increase of 4% in precision in comparison to thatobtained by using a TM image alone. A canopy elevationmodel combined with images from ALOS PALSAR,RADARSAT-2, and SPOT was used to classify vegetationsin the Alps and achieved 97.7% precision (Laurin et al.2013). The linkage of multispectral Li DAR and radar datayielded information on vegetation reflectance, height, andthe backscattering mechanism to allow for improvedmapping and characterization accuracy (Niculescu et al.2016). The highest accuracy of land use/land cover clas-sification was derived from multitemporal, multisensor,and multipolarization SAR satellite images (Huett et al.2016).The sole use of optical remote sensing data or micro-wave remote sensing data cannot achieve high accuracy forhigh-precision forest type recognition. However, the use ofa combination of data, such as optical remote sensing andmicrowave remote sensing data, offers complementaryinformation that can greatly improve accuracy, ispracticable, and leads to improved results. In this study, acombination of multiphase C-band data from polarizationRADARSAT-2 and SPOT5 optical images was used toanalyze different forest types and their polarization scat-tering features, spectral information, and phase character-istics in August and November 2013. The random forestclassification method was then used to classify the foresttypes of the Pangu experimental forest area.Materials and methodsStudy areaOur study area was at Pangu Forest Farm (Tahe ForestryBureau, Tahe County, Daxing0an Mountains, HeilongjiangProvince). Tahe County is located in the northwest of theDaxing0an Mountains in the northernmost part of China at123°200020 0–124°210400 0E and 52°160380 0–52°47040 0N(Fig. 1). The farm covers 1120.7 km2with elevations of800–1400 m. The climate is cool continental, with averageand maximum annual temperatures of - 2.4 and 47.2 °C,respectively. Annual precipitation ranges from 300 to450 mm and is mainly concentrated in July and August.Forest covers 88% of the total area. Dominant forest treespecies include Larix gmelinii, Pinus sylvestris, Betulaplatyphylla, Populus davidiana and Picea koraiensis.Remote sensing data sourcesPolarized RADARSAT-2 images in two phases and high-spatial-resolution SPOT5 images were used to identifyforest types. RADARSAT-2 is a high-resolution commer-cial radar satellite that carries a C-band sensor and waslaunched on 14December 2007 by the Canadian spaceagency and Mac Donald, Dettwiler and Associates Ltd.(MDA). The wavelength range of the C band is3.75–7.5 cm, and the orbital repeat cycle of RADARSAT-2 is 24 days. Additionally, POLSAR data were selectedfrom the HH, VV, HV, and VH polarimetry modes at twophases with the same orbital parameters, namely in the lushplant growth period from August 2013 and the leaf litterperiod from November 2013. The resolution was12 9 8 m. SPOT5 (French National Space Research Cen-ter) is an earth observation satellite in sun-synchronousorbit that was launched at the end of 2001. The maximumresolutions of the panchromatic and multispectral bands are2.5 and 10 m, respectively. The multispectral bandsinclude B1 (0.49–0.61 lm), B2 (0.49–0.61 lm), and B3(0.78–0.89 lm). Forest inventory data recorded duringearlier years were acquired for the study area, including thesub-compartment distribution. These data were used to1408 Y. Yu et al.123erify the results of the forest type classification madeusing remotely sensed data.Data preprocessingData preprocessing involved image filtering, terrain cor-rection, geometric correction, and registration of multi-phase SAR data and optical image data. First, SPOT5panchromatic and multispectral data were fused to acquirea fusion image at 2.5 m spatial resolution. This was fol-lowed by atmospheric correction, image mosaic, multi-look processing and SAR data filtering using Pol SARprosoftware, and by geometric correction and registration ontwo-phase SAR images based on the SPOT5 images afterorthographical correction. The polarization SAR imagewas resampled to 2.5 m by using the nearest-neighborresampling method to combine the optical images and SARdata.Method of classificationClassification systemA classification system was developed based on the presentsituation of the land use classification standard (Bu 2007),rules of forest resource survey in cities and counties inHeilongjiang Province, and in combination with remotesensing images and forest resource inventory data. Themajor forest types in the study area are mixed coniferousforests and mixed coniferous and broadleaved forests,namely B. platyphylla, P. sylvestris, L. gmelinii, and P.koraiensis forests. Mixed forests were not classifiedbecause the pixels might consist of identical features sincethe highest spatial resolution of SPOT5 images andresampled RADARSAT-2 images was 2.5 m. The foresttype classification system was designed for non-forests andB. platyphylla, P. sylvestris, L. gmelinii, and P. koraiensisforests based on the above factors.Fig. 1 Location of the study areaForest type identification by random forest classification combined with SPOT and… 1409123lassification by the random forest methodThe random forest method implements Breiman’s randomforest algorithm for classification (Breiman 2001), usesbootstrap samples of data and a decision tree. Successivedecision trees provide corresponding prediction results. Asimple majority vote is taken for the final prediction. Giventhat N samples are selected, the probability of each non-selected sample is (1 - 1/N)N. When the number of sam-ples (N) is sufficiently high, the probability converges to0.368 (1/e & 0.368), indicating that 37% of the samples donot appear in the training set to participate in the trainingmodel. The part that is not in the sample bag is termed Outof the Bag (OOB) and is used as a validation set to evaluatemodel performance. For each decision tree, the learningmachine produces an OOB bag for accurate estimates. ThisOOB is also used to obtain a running unbiased estimate ofthe classification error when trees are added to the forces toacquire estimates of variable importance. Variables withimportance values exceeding 0.01 are selected for classi-fication. Random forest classification displays high pre-diction accuracy and good tolerance to outliers and noise. Itis not easy to create an over-fitting phenomenon. Randomforest classification is a type of non-parameterized mod-eling tool with adaptive functions that are suitable to solveproblems resulting from a lack of prior knowledge and datawithout constraint conditions and rules. It effectively ana-lyzes interaction and non-linear relationship between dataand is used to handle substantial or multidimensional data.Feature extraction from RADARSAT-2 datafor classificationSeveral classification methods used by full-polarizationSAR data are based on decomposition theory. Thescattering characteristics of decomposed targets obtainedfrom polarization SAR data reflect features of differentobjects. Typically, target decomposition methods includecoherent and incoherent polarization decompositions. Theincoherent decomposition method is selected to decomposethe targets due to the complexity of natural targets. Featureextraction from RADARSAT-2 data for classification isdivided into three categories. The first category includes acovariance matrix, a coherent matrix, and eigenvaluesdirectly obtained from the original data. The second cate-gory is based on different decomposition methods andincludes several decomposition parameters (Cloude andPottier 1997; Krogager 2006; Freeman and Durden 1998;Huynen 1978; Holm and Barnes 1988; Yamaguchi et al.2006; Evans et al. 1988). For example, polarizationparameters of scattering entropy (H), scattering angle (a),and anti-entropy (A) are collected from a coherent scat-tering matrix based on the Cloude decomposition method.The third type includes the radar vegetation index (Linget al. 2009) and total power. Overall, 47 parameters areextracted from each RADARSAT-2 image (Table 1).Computational complexity increases if all polarizationparameters are used to identify forest types. The parametersare highly relevant. An increase in the number of param-eters used for classification increases noise to an extent thatforest types cannot be distinguished. Therefore, parametersin Table 1 should be eliminated first.The random forest model chooses variables by calcu-lating their importance such that the important variablereduces prediction ability and increases errors in the modelafter adding noise to these variables. The original OOBdata initially validate the model and increase its accuracy.A variable that adds noise to the OOB dataset is then usedto revalidate the random forest model to obtain a new levelof accuracy. The difference between the levels of accuracyTable 1 Parameters extracted from polarization decomposition of the RADARSAT-2 imageFeature Description ParameterOriginal data Covariance matrix [C]Coherent matrix [T]Eigenvalue k1, k2, k3Decomposition features Cloude decomposition H, A, aKrogager decomposition Krogager_ks, Krogager_kh, Krogager_kdFreeman decomposition Free_Vol, Free_Odd, Free_DblHuynen decomposition [T]_HuyHolm decomposition [T]_HolmYamaguchi decomposition Yamaguchi_Vol, Yamaguchi_Odd, Yamaguchi_DblVan Zyl decomposition Van Zyl_ Vol, Van Zyl_ Odd, Van Zyl _DblBarens decomposition [T]_BarRadar features Radar vegetation index RVITotal power Span1410 Y. Yu et al.123Forest type identification by random forest classificationcombined with SPOT and multitemporal SAR dataYing Yu1• Mingze Li1• Yu Fu1Received: 13 January 2017 / Accepted: 26 April 2017 / Published online: 14 November 2017Ó Northeast Forestry University and Springer-Verlag Gmb H Germany, part of Springer Nature 2017Abstract We developed a forest type classification tech-nology for the Daxing0an Mountains of northeast China usingmultisource remote sensing data. A SPOT-5 image and twotemporal images of RADARSAT-2 full-polarization SARwere used to identify forest types in the Pangu Forest Farmof the Daxing0an Mountains. Forest types were identifiedusing random forest (RF) classification with the followingdata combination types: SPOT-5 alone, SPOT-5 and SARimages in August or November, and SPOT-5 and two tem-poral SAR images. We identified many forest types using acombination of multitemporal SAR and SPOT-5 images,including Betula platyphylla, Larix gmelinii, Pinus sylvestrisand Picea koraiensis forests. The accuracy of classificationexceeded 88% and improved by 12% when compared to theclassification results obtained using SPOT data alone. RFclassification using a combination of multisource remotesensing data improved classification accuracy compared tothat achieved using single-source remote sensing data.Keywords Random forest classification Á Multitemporal ÁMultisource remote sensing data Á PolarizationdecompositionIntroductionAccurate classification of forest type is fundamental to thestudy of forest resources, forest dynamics, forest biomass,and carbon storage estimation. The use of remote sensingto aid forest type classification is increasingly important invirtually all aspects of forest research.Examples of remote sensing data used for forest typeclassification include TM and SPOT optical images. Thesetypes of imagery yield data to distinguish forest typesbased on spectral features and texture information that arereflected in a remotely sensed image. However, an objectcan be characterized by different spectra and differentobjects can have the same spectrum. This anomaly resultsfrom factors including weather, which affects opticalremote sensing images, and complicates classification offorest types through the use of remotely sensed data (Wangand Zhao 2005; Sun 2006). This situation is also observedin the Daxing0an Mountains where there are many foresttypes. It is more difficult to distinguish forest types in theseareas using only spectral characteristics. Microwaveremote sensing is a beneficial supplement to optical remotesensing because of its ability to perform day or nightimaging and all-weather imaging, penetrate clouds andrain, and generate increased information. PolarizationSynthetic Aperture Radar (POLSAR) data with full polar-ization has been used to produce information that isdirectly related to physical properties of natural media andbackscattering mechanisms including observational data,scattering matrix, covariance matrix, and correlationmatrix. Parameters extracted from these matrices throughdifferent polarization decomposition methods are appliedto the classification (Aghabalaei et al. 2016; Li et al. 2016;Lee et al. 1998; Cloude and Pottier 1997; Freeman andDurden 1998). Touzi et al. (2004) used the C band ofProject funding: The work was supported by the National NaturalScience Foundation of China (Nos. 31500518, 31500519, and31470640).The online version is available at http://www.springerlink.comCorresponding editor: Tao Xu.& Mingze Limingzelee@163.com1School of Forestry, Northeast Forestry University,Harbin 150040, People’s Republic of China123re calculated from the original OOB and new OOB withthe level of noise representing the importance of the cor-responding variable. An increase in the importance of themodel variable significantly decreases the accuracy calcu-lated using OOB data. The importance of the 47 parametersextracted from RADARSAT-2 images in August andNovember was calculated by the RF model. Highlyimportant parameters were chosen to identify forest types.The DEM in the study area was selected as secondary dataand used in forest type classification to reduce the influenceof topography. The DEM data were compiled from theASTER GDEM V2 data released by NASA in October2011at a resolution of 30 9 30 m2.Three schemes were used for classification by the RFmethod, namely the use of (1) SPOT-5 alone, (2) SPOT-5and SAR images in August or November, and (3) SPOT-5and two temporal SAR images.Separability of samplesThe ROI separability of training samples of differentobjects can be determined based on the Jeffries–Matusita(J–M) distance (Richards and Jia 1986). It ranges from 0 to2 and shows improved sample separability when the valueis closer to 2 (Ma et al. 2010). According to the forestresource inventory data and SPOT images, training sam-ples of different forest types were chosen to be evenlydistributed on images with obvious features: 200 uniformtraining samples for B. platyphylla, P. sylvestris, and L.gmelinii forests and 50 uniform training samples for P.koraiensis forest and non-forest.ResultsImportance of variables for classificationImportant parameters used for image classification acquiredfrom the POLSAR image in August and SPOT were C22,Van Zyl_vol, Yamaguchi_vol, T11huy, Freeman_vol, k1,Span, SPOT-5, and DEM; those used for image classificationacquired from the POLSAR image in November and SPOTwere H, Yamaguchi_vol, T11huy, T11, a, Van Zyl_odd,SPOT-5, DEM, Span, RVI, and T13huy_real; and those usedfor image classification on multitemporal POLSAR imagesand SPOT were RVI(11), Span(11), H(11), T11Holm(8),T11huy(8), SPOT-5, and DEM (superscripts of (8) and (11)refer to parameters extracted from the RADARSAT-2 imagein August and November, respectively). Figure 2 depicts theimportance of these variables.Separability calculationThe J–M distance was calculated for the three schemes(Fig. 3). The results with respect to the combination ofRADARSAT images from August and SPOT images, thecombination of RADARSAT images from November andSPOT images, and SPOT images alone indicate the fol-lowing: (1) insufficient differentiation of the B. platyphyllaforest because tree growth was more lush in August with arelatively similar scattering characteristic to coniferousforest. (2) Although scattering characteristics of coniferousand broadleaved forests are random and similar, the sepa-rability of training samples improved because the numbersFig. 2 Importance of the variables (Note A large decrease in accuracy indicates a more important variable)Forest type identification by random forest classification combined with SPOT and… 1411123of leaves of the broadleaved forest in November decreased.(3) The scattering and spectral characteristics of P. syl-vestris, L. gmelinii, and P. koraiensis forests were easilydistinguished without the effects of a broadleaved forest incontrast to when only SPOT images were chosen. Hence, afew polarization SAR image parameters were added forclassification. Highest separability of training samples wasobserved when combinations of spectral and scatteringcharacteristics from SPOT and RADARSAT images inAugust and November were used.Classification results and analysisForest type classification of the Pangu Forest Farm Clas-sification images using three schemes were shown inFig. 3 Separation of training samples by (J–M) distanceFig. 4 Forest type classification of the Pangu Forest Farm Classifi-cation images using a SPOT data (scheme 1); b RADARSAT-2images (August and SPOT data from scheme 2); c RADARSAT-2images from November and SPOT (scheme 2); and d SPOT andmultiphase RADARSAT-2 images from August and November(scheme 3)1412 Y. Yu et al.123Fig. 4. The classification precision of scheme 1 corre-sponded to 77%, indicating that the B. platyphylla forestwas accurately distinguished from coniferous forests,although it was confused with non-forests because a part ofthe open forest land of B. platyphylla was classified as anon-forest. With respect to the coniferous forests, L. gme-linii, P. koraiensis, and P. sylvestris forests were mixed to acertain extent. Therefore, classification using optical ima-ges alone is not sufficiently accurate, and thus, RADAR-SAT-2 images were added for forest type identification tosupplement optical images.The classification result precision was 80% when theSPOT and RADARSAT-2 images from August werecombined. The result evidently improved after addingdecomposed parameters from the RADARSAT-2 images.Although microwave data possess penetration characteris-tics and effectively overcome misclassification of the openforest land of B. platyphylla forest and non-forest, the dataare limited in their ability to improve the classificationcapacity because of random and complex scattering char-acteristics of vegetation in August.The classification result corresponded to 85% when theSPOT and RADARSAT-2 images from November werecombined, and was superior to the classification accuracyof using SPOT and RADARSAT-2 images from August. InNovember, all leaves in B. platyphylla forest fall and thisaids in distinguishing among L. gmelinii, P. koraiensis, andP. sylvestris forests, although the difference between B.platyphylla forest and non-forest was lower for increasedmicrowave scattering components from the trunk and sur-face and the multiple scattering between them.In scheme 3, the classification result precision was 88%when SPOT and RADARSAT-2 images from August andNovember were combined, an improvement in total accu-racy and precision because of the combination of themultiphase polarization characteristic parameters of theRADARSAT-2 images from August and November andoptical images. Multiphase features compensated for eachother with this combination.Discussion and conclusionsThe use of the RF classification method based on polar-ization information, spectral information, and phase char-acteristics reflected from multiphase microwave andoptical images provided more accurate forest classificationthan did any of these methods when applied individually.The use of only spectral information from SPOT5images confused coniferous forests because of their rela-tively close spectral characteristics with only 77% preci-sion. The addition of full-polarization SAR data fromAugust and November increased precision levels to 80 and85%, respectively. The maximum total precision was 88%for feature compensation following the introduction ofmultiphase RADARSAT-2 images.The complexity of the forest led to difficulties in featureextraction among different forest types. Texture informa-tion can be added since full-polarization SAR data wereused for forest type classification. Additionally, coherentinterference information from RADARSAT-2 images wasignored. In future studies, classification results can beimproved by combination with interference information.ReferencesAghabalaei A, Maghsoudi Y, Ebadi H (2016) Forest classificationusing extracted Pol SAR features from compact polarimetry data.Adv Space Res 57(9):1939–1950. https://doi.org/10.1016/j.asr.2016.02.007Borghys D, Yvinec Y, Perneel C, Pizurica A, Philips W (2006)Supervised feature-based classification of multi-channel SARimages. Pattern Recognit Lett 27(4):252–258Breiman L (2001) Random forests. Mach Learn 45(1):5–32Bu CGTZY (2007) Ministry of Land and Resources, P.R.CCloude SR, Pottier E (1997) An entropy based classificationscheme for land applications of polarimetric SAR. IEEE TransGeosci Remote Sens 35(1):68–78Evans DL, Farr TG, Van Zyl JJ, Zebker HA (1988) SAR polarimetry:analysis tools and applications. IEEE Trans Geosci Remote Sens26(6):774–789Freeman A, Durden SL (1998) A three-component scattering modelfor polarimetric SAR data. IEEE Trans Geosci Remote Sens36(3):963–973Furtado LF, Silva TSF, Novo EMLM (2016) Dual-season and full-polarimetric C band SAR assessment for vegetation mapping inthe Amazon varzea wetlands. Remote Sens Environ174:212–222. https://doi.org/10.1016/j.rse.2015.12.013Holm WA, Barnes RM (1988) On radar polarization mixed targetstate decomposition techniques. In: Radar, IEEE nationalconference—RADAR, pp 249–254Huett C, Koppe W, Miao Y, Bareth G (2016) Best accuracy landuse/land cover (LULC) classification to derive crop types usingmultitemporal, multisensor, and multi-polarization SAR satelliteimages. Remote Sens 8(8):684. https://doi.org/10.3390/rs8080684Huynen JR (1978) Phenomenological theory of radar targets.Electromagnetic scattering. Academic Press, New York,pp 653–712Kasapoglu NG, Anfinsen SN, Eltoft T (2012) Fusion of optical andmultifrequency polsar data for forest classification. Paperpresented at the 2012 IEEE international geoscience and remotesensing symposium. 22–27 JulyKrogager E (2006) Properties of the sphere, diplane, helix (targetscattering matrix) decomposition. Mol Ecol 15(11):3205–3217Laurin GV, Frate FD, Pasolli L, Notarnicola C, Guerriero L, ValentiniR (2013) Discrimination of vegetation types in alpine sites withALOS PALSAR-, RADARSAT-2-, and lidar-derived informa-tion. Int J Remote Sens 34(19):6898–6913Lee JS, Grunes MR, Ainsworth TL, Du LJ (1998) Unsupervisedclassification using polarimetric decomposition and the complexWishart classifier. IEEE Int Geosci Remote Sens4(5):2178–2180Forest type identification by random forest classification combined with SPOT and… 1413123