The Study Of Semisupervised Ensembled Support Vector Machines For Land Cover Classification

The coverage level of Earth’s surface can be reflected by land coverinformation, which is relative to production and living of human. For along time, the study of land cover changing has drawn much attention onresearch in the global environment community, and it has importantsignificance in terms of developing both social economy and ecologicalenvironment. To acquire accurate and timely information on land cover, itis important and urgent to develop effective approaches for land covermonitoring and mapping at large scale. Remote sensing technologies havethe capability of monitoring the Earth’s surface with different spatial,spectral and temporal resolutions. However, the current level ofclassification of remote sensing lags behind the development of remotesensing image acquisition technology. Therefore, the study of the newtheories and methods, which improves the remote sensing informationprocessing, has extremely important significance and wide applicationprospect. Support vector machines (SVM) is nowadays regarded as a newresearch focus on pattern recognition and machine learning fields. SVMhas the characteristics of simple structure, high adaptability, and globaloptimum and so on, and it can be used to solve the problems with high-dimension, non-linear, overfitting and uncetainty. SVM has beenwidely applied to land cover classification. There are some drawbacks forSVM method that should be improved, although the better classficationresults have been obtained for remote sensing data processing. The twokey shortcomings are that: firstly, the incorrectly selection of parametersaffects the generalization ability of SVM; secondly, the definition of aproper training set is another important issue in building SVMclassification system. The better results can not be obtained using SVMwhile the number of training samples set is far less than that of testsamples set, even if the SVM has the advantage of stronger generalizationability. Based on the problems proposed above, mainly contents andconclusions of this paper are listed as follows:1. A novel SVM parameters optimization method based onself-adaptive mutation particle swarm optimizer (AMPSO-SVM) isproposed to overcome the major drawbacks of the kernel functionselection and its parameters setting. The AMPSO algorithm, which isbased on the variance of the population’s fitness, can break away the localoptimum by the operation of self-adaptive mutation. Accordingly, veryhigh classification accuracy will be achieved with the best value of theparameters of SVM, which have been searched using AMPSO. In order toverify the validity of this AMPSO-SVM method, a remote sensing landuse/cover classification model is constructed using multi-spectral Landsat-5TM data. In particular, they are organized so as to test thesensitivity of the AMPSO-SVM model and that of the other referenceclassifiers used for comparison, i.e., Maximum likelihood classifier(MLC), SVM classifier and standard PSO algorithm for SVM parametersoptimization model (PSO-SVM). On an average, the AMPSO–SVMmodel yielded an overall accuracy of93.59%against83.92%formaximum likelihood classifer and outperformed PSO-SVM classifer interms of overall accuracy (by about2%). The obtained results clearlyconfirm the effectiveness and robustness of the AMPSO-SVM approachto the remote sensing land use/cover classification.2. A novel semisupervised SVM model that uses self-trainingapproach (PS3VM) is proposed to address the problem of remote sensingland cover classification. The major drawback for self-training algorithmis that classification errors can reinforce itself. If incorrect examples areincluded in the training set, then the final answer may be wrong. Hence, itis important to select the correct labeled points from the unlabeled set forself-training algorithm.The key characteristics of this approach are that1)the self-adaptive mutation particle swarm optimization algorithm isintroduced to get the optimum parameters that improve the generalizationperformance of the SVM classifier, and2) the Gustafson-Kessel fuzzyclustering algorithm is proposed for the selection of unlabeled points toreduce the impact of ineffective labels. The effectiveness of the proposed technique is evaluated firstly with samples from remote sensing data andthen by identifying different land covers regions in the remote sensingimagery. An overall accuracy of95.10%is achieved by the PS3VMmodel, which is2.04%and4.29%greater than that by S3VM (93.06%)and PSVM (90.81%), respectively. Experimental results show thataccuracy level is increased by applying this learning scheme, whichresults in the smallest generalization error compared with the otherschemes.3. Researchers have proposed separately semisupervised learningand ensemble learning methods in view of the small samples problemsthat improve the generalization performance of the SVM classifier.However, semisupervised learning and ensemble learning are indeedbeneficial to each other, and stronger learning machines can be generatedby classifier combination. Therefore, in this paper, a novel ensemblemodel with semisupervised SVM (EPS3VM) is proposed. PS3VMclassification model can produce several individual classifiers withdifferent performance, when it overcomes to the small samples problemleveraging large amounts of relatively inexpensive unlabeled data. Thenby the weighted voting method, these classifiers are combined so as toimprove the generalization ability of the classification model. Theeffectiveness of the proposed classification approach is demonstrated foridentifying different land cover regions in multispectral remote sensing imagery. In particular, the performance of the EPS3VM is compared withSVM and PS3VM in terms of classification accuracy and kappacoefficient. The EPS3VM model yielded an overall accuracy of96.88%against88.48%for SVM and outperformed PS3VM in terms of overallaccuracy (by about5%). The obtained results clearly confirm theeffectiveness and robustness of the EPS3VM approach to the remotesensing land cover classification.