Overview of deep learning algorithms for PolSAR image classification
Polarimetric Synthetic Aperture Radar(PolSAR)refers to an advanced microwave imaging system known for its robust observation capabilities in all-weather and all-day conditions.Therefore,the study of PolSAR image classification technology holds potential importance within the field.Recently,the emergence of deep learning technology has introduced novel research avenues for PolSAR image classification owing to its potent learning capabilities.In this context,the paper introduces a systematic review of deep learning-based PolSAR image classification methods.The paper begins by providing a concise review of the vital imaging mechanism of PolSAR,including polarization scattering representation methods,polarization target decomposition theories,and polarization data modeling.Then,this study presents detailed discussions on the application of deep learning techniques with different paradigms and network architectures for PolSAR image classification.The discussed works include a wide range of paradigms,from the traditional supervised and unsupervised approaches to the semi-supervised,self-supervised,and active learning approaches.Furthermore,the paper introduced various network architectures used in the PolSAR image classification domain,including Convolutional Neural Network(CNN),Recurrent Neural Network(RNN),Transformer,Restricted Boltzmann Machine(RBM),Deep Belief Network(DBN),Autoencoder(AE),Generative Adversarial Network(GAN),and Diffusion Model(DM).This paper highlights the potential benefits of integrating deep learning methodologies with polarimetric principles to augment the effectiveness of PolSAR image classification techniques,especially in dealing with complicated problems and enhancing model interpretability.On this basis,the paper summarizes the exploratory research on data-physics fusion in PolSAR image classification realm by dividing them into two classes including"polarimetric mechanism guided deep learning"and"deep learning simulated polarimetric mechanism".To the best of our knowledge,existing deep learning-based PolSAR image classification methods mostly use the former research idea.Therefore,based on the different polarimetric mechanisms guiding the deep learning methods,we further discuss the"polarimetric mechanism guided deep learning"methods by dividing them into two subclasses,i.e.,"methods based on data characteristics"and"methods based on scattering mechanism".Finally,the paper deliberates on existing challenges and future prospects in the realm of deep learning-based PolSAR image classification.One key challenge is the imperative need to integrate data-driven deep learning with PolSAR mechanism models to increase model performance and interpretability.Research in this domain is still in the early stages,providing opportunities for exploring novel patterns automatically extracted from PolSAR data.Addressing the limitations posed by the scarcity of labeled data remains a critical obstacle,with solutions like semi-supervised and self-supervised learning and the utilization of domain expertise to improve algorithm performance.The fusion of multimodal information from diverse remote sensing systems is another promising direction for observations.With the development of technology,designing large-scale PolSAR datasets and developing suitable large models tailored to SAR imagery are also vital areas for future studies.To conclude,deep learning methods have significantly improved the classification performance of PolSAR data in decades.Moreover,this paper provides a systematic overview of deep learning-based PolSAR image classification methods.The integration of data-driven deep learning with polarimetric mechanisms exhibits promising prospects.Challenges exist in applying these methods to real-world scenarios and in advancing research on data-physics fusion in PolSAR image classification realm.This paper identifies current challenges and future research directions,hoping to enhance the practical application of PolSAR technology.
万方数据
维普
