Polarimetric SAR image classification by Boosted Multiple-Kernel Extreme Learning Machines with polarimetric and spatial features

In this paper, we propose new approach: Boosted Multiple-Kernel Extreme Learning Machines (BMKELMs), a multiple kernel version of Kernel Extreme Learning Machine (KELM). We apply it to the classification of fully polarized SAR images using multiple polarimetric and spatial features. Compared with other conventional multiple kernel learning methods, BMKELMs exploit KELM with the boosting paradigm coming from ensemble learning (EL) to train multiple kernels. Additionally, different fusion strategies such as majority voting, weighted majority voting, MetaBoost, and ErrorPrune were used for selecting the classification result with the highest overall accuracy. To show the performance of BMKELMs against other state-of-the-art approaches, two L-band fully polarimetric airborne SAR images (Airborne Synthetic Aperture Radar (AIRSAR) data collected by NASA JPL over the Flevoland area of The Netherlands and Electromagnetics Institute Synthetic Aperture Radar (EMISAR) data collected by DLR over Foulum in Denmark) were considered. Experimental results indicate that the proposed technique achieves the highest classification accuracy values when dealing with multiple features, such as a combination of polarimetric coherency and multi-scale spatial features.