Target Recognition Of Floating Raft Aquaculture In PolSAR Images Based On Multi-feature Ensemble

Polarimetric Synthetic Aperture Radar(PolSAR)actively emits microwave signals and records the full-polarization information of scattering echoes of the target,which is conducive to the analysis of the target's scattering mechanism and is free of solar radiation and weather condition.Based on these,PolSAR has been successfully applied in target recognition and other fields.Floating raft aquaculture is an important type of dynamic monitoring of sea areas.Accurate recognition of the floating raft target is conducive to the rational use of aquaculture resources and the establishment of a healthy seawater ecological environment.Since that PolSAR can reflect the backscattering characteristics of the target,PolSAR images can effectively display areas of floating raft aquaculture and provide the data source for accurately recognizing the floating raft target.However,PolSAR images lack effective features and are affected by a large amount of irregular speckle noises.So these images have many isolated noise points,and the interior region of the floating raft target is incoherent and edges are not clear,which increases the difficulty of accurate recognition of the floating raft target.In order to overcome above problems,based on the characteristics of floating raft aquaculture,this paper designs a complete and effective process of floating raft aquaculture target recognition in PolSAR images.The main contributions are shown as follows:In order to effectively achieve multi-feature extraction,according to the scattering mechanism type and stripe imaging characteristics of the floating raft target,polarimetric features,texture features and contour features are extracted.Besides,the nonlocal multiple kernel fuzzy C-means algorithm is proposed for multi-feature integration.It designs a specific number of kernel functions with different kernel widths based on three types of features,and adaptively adjusts the weights of kernel functions,which can make full use of the beneficial kernel functions to improve the recognition result.Based on the data of floating raft aquaculture areas in the adjacent sea area of Changhai County,Liaoning Province,experimental results show that the proposed approach can accurately recognize floating raft targets,reduce isolated noise points,and suppress speckle noises.In order to effectively suppress speckle noises,a superpixel segmentation algorithm combining generalized local binary pattern and generalized statistical region merging is proposed.According to the multiplicative noise model,generalized local binary pattern is obtained to reduce the sensitivity of speckle noises.Then,the texture similarity criterion is added into the merging criterion of the generalized statistical region merging algorithm,which can obtain superpixels with texture consistency.After that,contour features are extracted to enrich the feature dimension of the data,and FCS algorithm is applied to conduct cluster to obtain the recognition result.Based on the data of floating raft aquaculture areas in the adjacent sea area of Changhai County,Liaoning Province,experimental results show that the proposed approach can accurately recognize floating raft targets for SAR images with different bands,resolutions,and regions,and it's universal and robust.In order to effectively implement feature optimization,classification and recognition,an adaptive nonlocal stacked sparse autoencoder network is proposed.According to the multiplicative noise model,the nonlocal mean algorithm is improved by adaptively optimizing neighborhood configurations of each pixel.Hence,adaptive non-local spatial information can be obtained,and is input to the network as the part of features.What's more,at the first layer of the network,original features and adaptive nonlocal spatial features are required to be close to each other,and more robust features can be obtained,whose effects would transfer to the rest of layers to optimize features.Softmax regression is used to implement classification and recognition in the end.Experiments were conducted on standard data and floating raft aquaculture data.The results show that the network can improve the classification and recognition accuracy,reduce isolated noise points,and obtain smoother and clearer class boundaries,which can verify the validity and practicality.