Land Cover Types Classification By Support Vector Machines Using Multi-temporal Polarimetric SAR Data

Remote sensing image classification can provide basis information for monitoring andevaluating resources of forestry and agriculture. Polarimetric SAR (POLSAR) imageclassification, as an important content of remote sensing image classification, has been paidmore attention in the international remote sensing research field. Most of the previousPOLSAR classification methods are based on single-temporal image, and the classificationaccuracy is not high in general. To address this problem, the land cover classification methodusing POLSAR image by support vector machine (SVM) has been studied in this paper. Thestudy site is located in Tahe County, Heilongjiang Province, China, which is near the borderbetween China and Russia. Two scenes of quad-Polarimetric Radarsat-2SAR images, oneSPOT-5image and ground true data were acquired for evaluating the accuracy.For the application of SVM for multi-temporal PoLSAR classification, firstly, wedeveloped the “Feature selection method based on exhaustive search strategies combined with testaccuracy”, and two temporals POLSAR data are used for land cover classification test. Theresults indicate that, the higher accuracy is obtained by the feature selection method developedin this paper than by the feature selection method based on knowledge and the non-featureselection method. The accuracys of these three method are84.96%，80.28%and65.68%.Secondly, the discernment ability of different temporal SAR data is evaluated by the sameclassification method based on SVM. The results indicate that the discernment ability of SARdata in October is higher than July，and the SAR data in July and October combined forclassification can offset their respective disadvantages and improve overall accuracy, theoverall accuracy of multi-temporal SAR data is the highest. Their accuracys are73.53%,57.08%and84.96%.The last, the result of classification method based on SVM is compared with that ofmaximum likelihood(ML). When the features’ probability obey the gaussian distribution, it shows that the accuracys of SVM and ML are84.96%and72.65%, respectively. Although thefeatures’ probability do not obey the gaussian distribution, the accuracy of SVM is higher thanML too, the accuracys are65.86%and63.74%, respectively.