Land Cover Classification Based On Model Clustering And SVM

In this study,Select2000,2005,2010three years of MODIS EVI data which are provided every16days at500-meter spatial resolution as a gridded level-3product in the Sinusoidal projection. A novel hybrid Mclust and SVM for land cover classifica-tion:a case study in Loess Plateau which vegetation and climate are very sensitively. The main steps include:First, data preprocessing, including scaled and KPCA transfor-mation. Second,random sample20000+points, Gaussian mixture Modeling for Model-Based Clustering(Mclust) was used on these points. In order to have a representative sample, repeat20times. Third, employing the SVM method to extend the clustering results of random samples to the entire region, then combining with auxiliary data and other information(e.g. Google Earth) for class definition, merge, verify the accuracy and so on. Finally, based on land cover2010data, extracting different land cover types char-acteristic curve, combining with LibSVM model for land cover automatic classification in2000and2005. Then according to the three land cover data, a brief description of the different periods and regions of the Loess Plateau land cover change and conversion types during the past10years.The main conclusions are as follows:(1) Establish a novel hybrid Mclust and SVM for land cover classification, and the validation results show that the overall accuracy of87%, Kappa coefficient was0.75in Loess Plateau.(2) Loess Plateau has a strong zonal vegetation:overall, scarcity of forest resources; natural vegetation including broadleaf deciduous forest, forest-grass, grass and desert grass from southeast to northwest; crops including Weihe plain two-season crops, main-ly winter wheat and summer maize; Fen River plain includes single-season rice in south-west and the Northeast dominated by grain and cotton; Mainly single-season wheat in Hetao plain and Ningxia.In2010, the area of forest, grass, cropland, bare and water were97670.81km2,399434.18km2,105129.64km2,14764.46km2,3000.94km2, respectively.(3) Overall, the main conversion types include grass converted to cropland, forest and bare; and cropland into grass, forest and bare in Loess Plateau. Nearly10years of forest greatest increases, reaching105.411%, while bare fell the most, was-48.034%. In addition, according to two periods (separated by5-year) conversion results can obtain two main mixed mode that cropland converted to grass, then into forest, and bare converted to grass then into forest.(4) A brief analysis of the driving forces of the land cover change, and the results showed that:on the one hand, human factors including frequent fanning, grain to green, and wasteland suitable for afforestation; on the other hand,rainfall on land cover change played a key role.The main highlights:Integration of the more popular model-based clustering (Mclust) and support vector machine (SVM), the algorithm not only improve the classification accuracy, but also weakened the different people in the land use/cover differences.Based on the results of land cover classification in2010, extracting different land cover types characteristic curve. Then use LibSVM model for land cover auto-matic classification. Not only saves manpower and resources, causes confusion by different people or at different times of operation, also better indication of land cover change between years.A case study in Loess Plateau that complex and diverse bioclimatic zone. On the one hand, from the reality verify the feasibility of the mixed method, as well as the foundation for the relationship between people and environment in Loess Plateau further research.