Research On Polarimetric SAR Image Classification With Limited Sample Size

Synthetic Aperture Radar(SAR)is an important technology to observe the Earth’s surface through actively transmitting and receiving microwave signals,This imaging model can break through the constraints of unfavorable meteorological observation conditions such as clouds and rain,so that all-day and all-weather continuous observation is realized.Additionally,the microwave signals can penetrate the observation target to a certain extent,and obtain a rich description of scattering characteristics and structure information which is different from optical remote sensing in the data acquisition capability and target structure information sensitivity characteristics.Benefitting from its multi-polarization characteristics,polarimetric SAR can obtain more detailed descriptive information of target backscattering,possessing the potential of applying g in the broad application in various fields such as disaster monitoring,environmental monitoring,and forest surveys.As an important prerequisite for the boosted understanding and interpretation of polarimetric SAR images,ground targets classification has always been a research hotspot in the field of polarimetric SAR.Especially,with the ongoing improvement of polarimetric SAR theoretical research and computer science,the classification framework based on deep learning has brought breakthrough progress in polarimetric SAR image classification research in various scenarios,which has now become the important study direction in this field.However,the aforementioned research paradigm often requires a large number of high-quality labeled samples as input or assistance.In practical classification tasks,as the sample collections are easily affected by various factors such as rich target varieties and complex imaging environments,it is often difficult to obtain a sufficient number of labeled samples to ensure the effective performance and generalization ability of the classification model.As a result,how to obtain highly accurate classification results when the number of labeled samples is limited,that is,the problem of small samples,still remains the hotspot as well as a tough difficult point in the current field of polarimetric SAR classification research.The problem of polarimetric SAR image classification under the condition of small samples can be further categorized into two main aspects in terms of data representation and model learning.To start with,with respect to data labeling,when considering the diverse spatial distribution of targets,it is,in most cases,difficult to obtain a satisfactory description of multi-type objects in a complex surface environment with the usage of limited sample data.Moreover,in terms of model learning,taking into consideration of the multi-polarization and multi-dimensional data characteristics of polarimetric SAR data,the lack of sample data is easy to further trigger the high-dimensional small sample problem.Here,the high-dimensional small sample problem means that with the increase of feature dimensions,the computational complexity of classification increases significantly,which in turn makes the model require a surge in sample amount.This snowball-like effect eventually leads to low classification accuracy and poor model stability.In terms of the above analysis,this paper conducts research on polarimetric SAR image classification under the condition of small samples from three main aspects,including improving the ability of samples to represent different classes,boosting the training strategy of classification models,and improving the training process of largescale classification and mapping.The specific research work of the present study can be summarized as follows:(1)In terms of the improvement of the data representative capability of target objects,this study proposes a feature selection method for the polarimetric SAR image classification task based on a multi-scale two-dimensional structural similarity measure.Taking into combined consideration of both polarimetric and spatial features,the proposed method achieves the adaptive feature selection of high-dimensional features of polarimetric SAR images by measuring the "feature-category" intra-class structural similarity and inter-class structural difference at different spatial scales.The method can effectively improve the representation capability of the feature set to the target objects while reducing the data dimension of the classification feature,thereby helping to improve the classification accuracy of polarimetric SAR images.(2)With respect to the improvement of the training strategy of the polarimetric SAR image classification model based on combined polarimetric-spatial features,the present study develops a polarimetric SAR image classification framework based on Reinforcement Learning(RL).The proposed framework combines the powerful feature learning ability of the Fully Connected Convolution Neural Network(FCN)and the experienced "explore-utilization" training strategy of reinforcement learning,which can yield experience data(not original labeled samples)through model bootstrapping during the training process and improve the model learning strategy through experience replay.This unique way of generating and utilizing empirical data,to a certain extent,increases the utilization of original labeled data,thereby helping to improve the accuracy of image classification under the condition of small samples.In addition,for different classification scenarios,this study further explores two model training strategies within the framework,that is,one strategy that combines supervised learning pre-training,and the other training from scratch.Based on these two strategies,two classification methods are accordingly realized to apply the Polarimetric SAR image classification task in different scenarios.(3)In the aspect of large-scale polarimetric SAR image classification and mapping with combined polarimetric-space-time features,a large-scale polarimetric SAR image classification and mapping framework is proposed in this study based on small-scale model pre-training and model transfer.First,the framework proposes a temporal attention classification method that takes into account the temporal phenological features of the target objects and the correlation of polarized channels,on the basis of combining the multi-temporal features of the objects,so as to improve the model’s attention to the key temporal features of the target.Then,the classification accuracy of polarimetric SAR images under the condition of small samples is improved;Second,built on the first steps,for the problem of small samples in large-scale classification mapping,that is,a small sample rate also means a large number of actual sample collection,and The spatial distribution of the samples needs to satisfy a relatively regular spatial distribution.The large-scale classification mapping method based on the migration of the small-scale pre-training model is studied,and the accuracy of the largescale classification mapping of the framework is verified in the comparison with the optical reference ground truth data.