Polarimetric SAR image classification based on selective ensemble learning of sparse representation

This paper presents a sparse representation (SR)-based selective ensemble learning method for Polarimetric SAR image classification. Sparse representation uses the least dictionary atoms which come from a structured dictionary to represent the data, however, different training samples will lead to different options of the selected atoms and the corresponding coefficients, which will lead different subsequent classification results. Ensemble learning can be adopted to solve the issue, which uses several different learners to acquire a set of results that are integrated to get the final result. But, the result of each learner isn't all good. Therefore, the paper introduces selective ensemble learning which excludes the learners whose weights are smaller than the pre-set threshold. Experiments are conducted on the PolSAR image data of San Francisco test area to verify the performance of the proposed method.