Evaluation of Multiple Classifier Combination Techniques for Land Cover Classification Using Multisource Remote Sensing Data

Use of multisource remote sensing data, particularly Synthetic Aperture Radar (SAR) and optical images, can improve performance of land cover classification since many types of information are involved in the classification process. Recently, the multiple classification systems (MCS) or ensemble classifiers has been developed and increasingly used for classifying remote sensing data. It is considered as a promising approach to increase the classification accuracy. In this paper, different classification combination methods were carried out and evaluated for classifying land cover features in New South Wales, Australia using various integrations of SAR (ALOS/PALSAR, ENVISAT/ASAR) and optical (Landsat 5 TM+) satellite images and their derivative products such as textural information, Normalized Difference Vegetation Index (NDVI). Three classifiers were applied for classification, including Artificial Neural Network (ANN), Support Vector Machine and Self-Organizing Map (SOM). The outputs of these classifiers were then fused using various combination schemes such as majority voting, sum, evidence reasoning (Dempster-Shafer) theory. The other approach involve using other well known MCS techniques, namely, bagging and boosting algorithms were also carried out for each of the classifier. Results of the study illustrated the advantages of the multiple classification combination approach for land cover classifications, especially in conjunction with multisource remote sensing data. In most of cases, the multiple classifier system outperformed the single classifier and gave a noticeable improvement in the classification accuracy. The experiments also revealed that the multisource datasets always gave better accuracy than that obtained by the single-source datasets.