Object-oriented land cover classification of HJ-1B CCD image through multiple classifier fusion

Recently, classifier fusion has shown great potential to increase classification accuracy beyond the level reached by individual classifiers. In this work, we design a strategy to fuse several classifiers aim to improve the land cover classification accuracy effectively. Use the multi-spectral remote sensing image of HJ-1B CCD as the data source, the correlations of different raster bands are analyzed and different features are extracted for multi-resolution image segmentation and classification, such as the NDVI, NDWI. And then, several classifiers are adopted for object-oriented land cover classification, including multiple support vector machine (SVM) with the core of the radial based function (RBF), SVM with the core of linear function, Neural network (BP), decision tree of rough set, random forest, and K nearest neighbor. Finally, classification results from different classifiers are fused to improve the reliability and robustness of the results. A case study using HJ-1B multispectral images located in ShanXi province has proved the effectiveness of the proposed method.