Decision fusion of classifiers for multifrequency PolSAR and optical data classification

Forest detection and classification in tropical regions is very important for climate change research. Combining available data from different sensors is widely used in remote sensing to improve detection and classification performance. In this study, a decision fusion strategy is proposed to integrate optical and multifrequency PolSAR data for classification of rural areas including forest. Developed decision fusion strategy was validated with testing and validation samples which were manually selected from the high resolution satellite imagery. A total of three different sensor-originated scenes acquired on May 2010 in the Northwest of Tanzania were used in forest detection and classification experiments. The results show that combining classifiers for combinations of different sensor-originated features improves classification results for detailed class categories. Features which are properly modeled with the same statistical distribution are grouped and processed together. Classification results are weighted by using a reliability measure which is derived from confusion matrix of validation set. Therefore proposed decision fusion strategy improves the performance of parametric classifiers for some cases.