Research On PolSAR Image Classification Method Based On Recurrent Network

Polararimetric synthetic aperture radar is a high-precision imaging radar that can obtain rich target and land cover information.Polararimetric synthetic aperture radar not only has high imaging resolution,but also has the characteristics of all-weather ground observation,which is not disturbed by light and climate conditions during operation,and can even penetrate masks to obtain information about its coverage.These advantages of synthetic aperture radar make it widely used in agriculture,environment,geology and military fields.Traditional Pol SAR image classification methods usually ignore the spatial information between the pixel points of Pol SAR images,and the classification results often fail to meet the expectations.The main work of this paper is the research of Pol SAR image classification method based on recurrent network,which constructs Pol SAR spatial sequence based on long and short-term memory network and applies it to Pol SAR image classification,and the main work is as follows.(1)To address the problem of spatial information loss caused by traditional Pol SAR image classification methods,this paper proposes a Pol SAR image classification method based on attention mechanism for deep sequence networks.The method increases the spatial information between pixels through spatial sequences.First,the LSTM network converts time series into spatial series to extract spatial features.Then,an LSTM-based spatial enhancement strategy is proposed to enhance the relationship between the spatial information of pixels.Finally,to avoid the feature selection process,an attention mechanism is introduced in the LSTM network to select the important information and improve the classification performance.The experimental results show that the proposed spatially enhanced LSTM network can increase the spatial information between pixel points and achieve better classification results even when a small number of samples are used.(2)In order to compensate for the weak feature extraction ability of the traditional Pol SAR image classification method,a multi-scale spatially enhanced LSTM Pol SAR image classification method based on the spatial attention mechanism is proposed in this paper.The method consists of an automatic feature extractor,selector and classifier,which does not need to extract and select features to achieve end-to-end classification.A multi-scale fusion of high-level features is constructed between different input sequence data to increase the multi-scale spatial features and reduce the effect of input sequence data in different order.Finally,a new loss function is defined to reduce the risk of network overfitting and improve the classification results.The experimental results show that the proposed method can compensate for the information loss problem caused by the feature extraction process,and adding a spatial attention mechanism to the method can improve the classification accuracy of the model.(3)To address the problem that the traditional Pol SAR image classification method does not consider the physical characteristics of Pol SAR images,this paper proposes a Pol SAR classification method based on the physical guidance mechanism.The method firstly performs Freeman-Durden decomposition on the covariance matrix of Pol SAR data,and decomposes the Pol SAR data into body scattering Pv,surface scattering Ps and dihedral angular scattering Pd,three scattering mechanisms.Next,these three scattering mechanisms are input to the depth autoencoder for training.Then,the spatially enhanced LSTM network is trained using the selected data with labels.Finally,the two models are migrated by transfer learning,and the extracted physical and high-level features are fused to obtain the final results by Softmax for classification.The experimental results show that the physically guided Pol SAR image classification method based on physical guidance can improve the classification accuracy.