Study On PolSAR Image Classification Method With Superpixel Convolutional Neural Network

Synthetic Aperture Radar(SAR)is an active microwave Radar imaging system with all-weather and all-day imaging.The system has certain penetration and can obtain high-precision surface scattering information in a large area.It plays an important role in military and civil fields.Polarimetric SAR(PolSAR)can transmit and receive polarimetric echoes in different combinations,which provides rich backscattering information for SAR image interpretation.The expression of PolSAR image information is complex and diverse,and the effective expression of the scattering mechanism of ground truth and efficient extraction of polarimetric characteristics need to be further studied,which poses a higher challenge for the interpretation of PolSAR image.On the basis of superpixel segmentation and focusing on the interpretation of PolSAR images,this paper conducted a study on PolSAR image classification using the superpixels obtained by Minimum Spanning Tree(MST)with a small number of labeled samples:(1)The pixel-level classification method ignores the spatial relations and speckle noise of PolSAR image.Also,the classification strategy of per-pixel classification fails to take into account the homogeneity within the same category.Therefore,the MST superpixel classification method based on the Convolutional Neural Network(SPCNN)is proposed.SPCNN takes superpixel as the basic classification unit,which can effectively reduce the number of units to be divided when classifying the whole image.Specifically,the method combines pixels with polarization similarity through clustering strategy by MST to obtain irregular superpixel blocks with spatial consistency.The normalized patch is input into the CNN model by the method of superpixel regularization.The experimental results show that compared with the traditional method,the SPCNN algorithm can obtain the classification results with clearer plot boundaries and higher homogeneity within the plot,and also shows great advantages in the average classification accuracy of most ground features.(2)For the performance of SPCNN is poor in the case of a small number of labeled samples.A superpixel classification method based on convolutioanl deep belief network(SPCDBN)is developed from the perspective of semi-supervised learning in this paper.The SPCDBN model firstly pretrains with a certain amount of unlabeled data,and then initializations of the pre-trained model parameters into the classification model.Finally,ideal classification results can be obtained by fine-tuning with a few labeled samples.Experimental results show that SPCDBN can effectively increase the robustness of discriminant network to random initialization parameters by pre-training unlabeled data.At the same time,a proper amount of unlabeled pre-training data can accelerate the convergence of the model to the optimal value,and only a small number of iterations can achieve high classification accuracy.Compared with the other classification methods,the SPCDBN proposed in this paper can perform well in the average accuracy of most ground objects when the training samples are small.(3)To solve the problem that the classification performance of classification model with single scale is easily affected by the parameter of segmentation scale.In this paper,a back fusion strategy of multiscale superpixel classification(BFMSSP)method based is developed.It overcomes the difficulty of segmentation scale selection in superpixel classification.Specifically,BFMSSP achieves the classification task by two steps: First,the whole PolSAR image is segmented according to a certain scale value to obtain multi-scale segmentation results matching the normalized patch size,which have hierarchical structure connections.In the segmentation results,the normalized patches corresponding to the three super-pixel segmentation scales were selected as the feature inputs of small,medium and large scales to CNN respectively,and the classification results at different scales were fused on the label map by per-pixel voting.Experimental results show that BFMSSP algorithm can provide more accurate and complete classification results.Meanwhile,the proposed algorithm makes full use of multi-scale superpixel information and further improves the robustness of the classification model.