| Polarimetric synthetic aperture radar(PolSAR),as one of the most commonly used means in radar imaging technology,is widely used in various fields,among which the effective classification of PolSAR terrain is a research hotspot in the field of remote sensing.In recent years,with the rapid development of deep learning,it has achieved excellent performance in solving the problem of PolSAR terrain classification.The PolSAR terrain classification is a small sample classification task,and the traditional deep learning needs to obtain feature information from a large number of labeled samples to optimize the network model,so how to fully extract the feature information of PolSAR data under the condition of small samples is a key problem to be solved.Although many scholars have discussed this problem,they usually use the way of sample expansion to expand the dataset and fail to make full use of the existing samples.In view of the fact that the UNet has a good ability to segment small sample datasets,the paper proposes three PolSAR classification methods based on UNet according to the unique attributes of PolSAR data.The specific research results are as follows:(1)Considering the unique data attributes of PolSAR images and the influence of deeper feature on model performance,a PolSAR terrain classification method based on UNet is proposed.Firstly,considering the unique data structure of PolSAR,depthwise separable convolution is introduced to extract the spatial features and channel features of data separately.The use of depthwise separable convolution can not only reduce the extraction of redundant features,but also reduce the complexity of the model.Secondly,residual structure is introduced to extract deeper feature information.Finally,depthwise separable residual network model based on UNet(SR-Unet)is proposed by combining UNet with depthwise separable convolution and residual network,and the classification task of PolSAR terrain is completed.The experimental results show that the performance of this method is better than other comparison methods.(2)On the basis of SR-Unet and considering the small sample characteristics of PolSAR dataset,a multiscale residual UNet model based on attention mechanism(AMRUnet)is proposed.Firstly,in order to make full use of labeled samples to obtain richer feature information,a multiscale residual structure is proposed,in which the feature information of different receptive fields is extracted by setting convolution kernels of different sizes.Secondly,the multiscale residual structure is improved by using the idea of dense connection,and the improved multiscale residual structure can effectively improve the utilization rate and transmission efficiency of feature information.After that,the spatial attention mechanism of receptive field fusion is introduced to assign different weights to different features,which further enhances the learning ability of the network model for important features.The experimental results clearly show that this method can not only effectively improve the classification accuracy and of PolSAR terrain,but also improve the efficiency.(3)Considering the importance of phase information and multiscale features to the PolSAR terrain classification,an improved complex-valued dual-channel UNet(CV-DUnet)is proposed.Firstly,a dual-channel downsampling module is designed to extract multiscale features of PolSAR data.Secondly,an improved jump connection is introduced to eliminate the possible semantic differences between high-level and low-level features.Finally,the network model is extended to the complex-valued domain to make full use of the phase information of PolSAR data.The experimental results show that the network model improves the classification accuracy of PolSAR terrain to some extent. |
