Deep Representation Learning And Dynamic Scaled Strategy Based PolSAR Image Classification

Polarimetric Synthetic Aperture Radar(Pol SAR)is an active imaging tool,whose data acquisition process is not affected by season,weather,light,and other factors.The acquisition of massive Pol SAR data makes it possible to better analyze the types of land-covers in specific areas.With the development of technology,Pol SAR image classification has gradually changed from the early classification mode based on statistical characteristics and manually designed features to the research of deep learning methods having adaptive learning ability.On the one hand,the data representation ability of deep learning is becoming stronger,so that the classification performance of Pol SAR images has been comprehensively improved.The powerful ability of the deep learning models depend on the guidance of a large amount of annotation data.For different machine vision topics,many large-scaled annotation data sets using optical images,such as Image Net,CIFAR10,COCO,have been proposed,greatly promoting the progress of image classification,target recognition,video motion detection and other vision fields.For Pol SAR data,it is difficult to label because it is different from the general optical image,which limits the representation ability of deep learning technology on Pol SAR data to some extent,and restricts the improvement of classification performance.On the other hand,Due to the geometric characteristics of Pol SAR imaging,such as scale distortion,top-bottom displacement,and perspective shrinkage,the ground objects at different positions will be different due to their own height,relative distance from the radar,and incident angle.The resulting Pol SAR image will also be different in scale,that is,the scale of ground objects changes dynamically with the spatial position.Therefore,how to extract effective representation information referring to deep learning technology in the case of limited Pol SAR labeling data and dynamic changes in the scale of polarimetric objects,so as to obtain satisfying Pol SAR image classification results,is also currently in need of research.Based on the above actual demand,this thesis proposes a series of models to improve the performance of Pol SAR image classification.The main contributions are as follows:1.Aiming at the randomness of scattering target echo direction in Pol SAR images,a polarimetric multi-path convolution neural network(Pol MPCNN)is proposed for Pol SAR image classification.First,in order to adapt to the larger and smaller target in Pol SAR images,a two-scale sampling is designed as the input structure of the Pol MPCNN.Next,according to the data form of two-scale sampling,the polarization rotation kernels(PRKs)are designed,and the polarimetric convolutional neural network(Pol CNN)is constructed based on it,which can adaptively learn polarization rotation angles,and is the base network of the Pol MPCNN.In order to learn the polarization rotation angles related to different type of land-covers directly,the Pol CNN is further extended to a multi-path structure,and the final model Pol MPCNN is obtained.In order to get more attention to the samples that are hard to train,the proposed stage-wise training algorithm is used in the training stage.The experiments on real Pol SAR data show that the proposed model can obtain the best results at a very low sampling rate compared with the contrast methods.2.Aiming at the problem that the difficult data annotation of Pol SAR images is difficult,so that the large amount of Pol SAR data available is hard be used for the training deep learning models,a semi-supervised patch-level Pol SAR image classification model,two-staged contrastive learning and sub-patch attention based network(TCSPANet),is proposed combining contrastive learning and Transformer.Firstly,a two-stage contrastive learning network(TCNet)is designed,which can learn the representation information of Pol SAR image patches and the discrimination and contrast of image blocks with real ground truth information under semi-supervised condition.Then,using Transformer,a sub-patch attention encoder(SPAE)is constructed,and the complexity of the land-covers of patch samples is analyzed by modeling the context inside image patches.In order to overcome the difficulty of pixel-level labeling of Pol SAR and fully train the proposed network,this chapter proposes two patch-level datasets,namely,the unsupervised multi-scaled patch-level dataset(Us Ms PD)and the semi-supervised multi-scaled patch-level dataset(Ss Ms PD),which contain not only the labeled image patches in Pol SAR images,but also the patches labeled with patch-level categories,as well as the automatically generated patch samples include multiple land-covers,so that the TCSPANet can learn the characterization information of Pol SAR patch samples composed of different scales and different ground objects that is conducive to image classification.In the prediction stage,the classification algorithm of "splitting or classifying" is used to classify the complete Pol SAR image and finally achieve the non-overlapping and scale adaptive patch-level classification effect,which not only improves the classification efficiency,but also makes the patch size better matches the scale of the ground object.3.In the Pol SAR classification method based on patch as the prediction unit,patch samples with a specific scale cannot flexibly cope with the problem of scale diversity of ground objects,a dynamic patch-level input network(DPINet)is proposed for Pol SAR image classification,which is more flexible in selecting the sample size of a classification patch.Firstly,a patch analysis network(PANet)is established to analyze whether a patch sample can be assigned a patch-level class by learning the percentage score for each land-cover category.For unclassifiable patches,cropping is needed to improve the score of main land-cover.For this reason,a patch crop network(PCNet)is constructed,and the appropriate patch cropping scale is deduced by reinforcement learning.Running the PANet and PCNet alternately,the scale of a patch sample is selected dynamically until the patch samples can be classified.The experimental results show that the proposed model can achieve good regional consistency and better detail retention.4.Aiming at the problem of parameter redundancy in dynamic patch-level input classification network with fixed model structure,a novel patch-level classification model-lightweight dynamic patch-level input network YORAP is proposed.For an initial sample patch,the proposed lightweight patch analysis network(LPANet)is used to analyze the separability for the patch with lower a mount of parameters and computation.Next,in order to improve the separability of block samples,a scale guided patch crop network(SGPCNet)is established,which can avoid excessive cropping while obtaining the cropping scale that makes the land-cover composition of the patch as simple as possible.Alternately perform LPANet and SGPCNet to obtain the final patch sample that can be classified,and use an additional patch classifier to identify its patch-level categoriy.During the whole prediction process,the patch scale and model parameters are dynamically changing.Only for classifiable image patches will all the network structures be used,so the proposed model can be summarized as you only recognize appropriate patch(YORAP).The experimental results show that the proposed model can obtain more reasonable patch-level classification results and competitive accuracy under the condition of less network parameters.