A decision Fusion method for land cover classification using Multi-sensor data

In recent years, decision fusion techniques have been widely applied in many studies to combine information from different sensor data to achieve higher accuracy in information extraction than that could be achieved by the use of single sensor data alone. So far, most of the decision fusion techniques developed for the remote sensing applications have the drawback of assuming conditional independence between the classification results and integrates only two sensor data. The proposed Multi-sensor Fusion of Correlated Probabilities (MFCP) algorithm is rooted in the Baye's theorem and is the extended version of the Fusion of Correlated Probabilities (FCP) method by accounting for the cross conditional dependence between three multi-sensor information sources. At first, this dependence (or correlation) structure is efficiently formulated from the binary classification results of SVM by considering three sensor data and then finally used as weighting parameters in the proposed novel MFCP based fusion method. The proposed MFCP based fusion method is assessed in the multi-sensor land cover classification over the Beijing area. The data sets used consisted of two optical data sets (i.e., Landsat TM and CBERS) and one multi-temporal Synthetic Aperture Radar (SAR) data set. We evaluated and validated our proposed methodology by comparing it with four existing fusion methods, i.e., Max rule based decision fusion method, Consensus theory based decision fusion method, SVM based decision fusion method and Tau model based decision fusion method. To further validate the effectiveness of the proposed MFCP based fusion method, we performed a significance test for kappa's between different fusion results. The experimental results demonstrated that the proposed MFCP based fusion method outperformed all the compared fusion methods that do not take into account the cross conditional dependence between classifications, in terms of overall accuracy, kappa and class wise accuracies. ? 2016 IEEE.EI145-14