High-Resolution PolSAR Scene Classification With Pretrained Deep Convnets and Manifold Polarimetric Parameters

How to jointly use spatial and polarimetric information in PolSAR analysis has long been an open question. Benefiting from advanced architectures and large visual databases, deep convolutional neural networks or deep convnets (DCNNs) can generate high-level spatial features and achieve state-of-the-art performance in image analyses. However, because PolSAR data are not only multiband but also complex valued, these models cannot be easily borrowed to process them. In light of this problem, we develop a new data set to explore the abilities and potentials of DCNN on PolSAR scene classification. We observe that these models learn fixed semantic information in each layer and adapt to a different data type via changing middle-level filters. Instead of detecting colorful patterns, filters for PolSAR data tend to generate features in separate colors, which may naturally enable the network to differentiate polarimetric mechanisms. Therefore, an ensemble transfer learning framework is proposed to incorporate manifold polarimetric decompositions into a DCNN without throwing away the prelearned spatial analytic ability. Different polarimetric parameters can reflect polarization mechanisms in diverse aspects and introduce new discriminative features to enhance the object recognition. The framework achieves 99.5% validation accuracy and may benefit PolSAR applications in a wide spectrum of fields.