Polarimetric SAR Image Classification by Multitask Sparse Representation Learning

Classification is an important and difficult problem in Polarimetric SAR (POLSAR) image processing. Most existing classification methods combine multiple features (scattering parameters or statistical distribution) to improve the performance. However, based on the observation that various regions have different characteristics due to the different scattering mechanism, which implies that different features should be used for certain pixels rather than using the combination of various features for the whole image, so that simple combinations will result in numerous error classifications. In this paper, a novel POLSAR classification method based on multitask learning with multiple features is proposed. Firstly, different types of features are extracted, and then POLSAR classification problem is formulated as a multitask joint sparse representation learning problem. The strength of different features are employed by using of a joint sparse norm. Finally, experimental results on real POLSAR data show that our method outperforms several state-of-the-art algorithms.