Comparison Study of Multitemporal PolSAR Classification Using Convolutional Neural Networks

Target classification is a main application of polarimetric synthetic aperture radar (PolSAR). With the PolSAR images of the same region observed on various dates, multitemporal classification plays an important role in crop discrimination and growth monitoring. However, due to the crop growth change, classification methods with good generalization are imperative. Recently, advanced deep learning techniques have achieved breakthroughs in optical image processing. Among them, convolutional neural networks (CNNs) can greatly improve accuracies in classification tasks. Because their multilayer architectures can extract the abstract feature with strong robustness to characterize the kinds' difference. In this vein, the multilayer architectures may be able to lead to the higher accuracies, as well as the better generalization in multitemporal PolSAR classification. So, this work aims to investigate multitemporal PolSAR classification using five CNNs of AlexNet, VGG16, Inception, ResNet50, and MobileNet. In detail, four roll-invariant features and two hidden features in the rotation domain are selected as inputs. Comparison experiment based on four temporal UAVSAR data with seven land covers validates the efficiency of these CNNs. For the train-used temporal, Inception can achieve the higher overall accuracies. While, for the train-not-used temporal, ResNet50 is with the better generalization.