Research On SAR Image Scene Classification Method Based On Deep Metric Learning

As an active detection system with all-weather,timeless,high-resolution and long distance,synthetic aperture radar(SAR)has been widely utilized in various military and civilian applications.It mainly includes automatic target recognition,crop growth monitoring,vegetation monitoring,natural hazard risk assessment and marine monitoring,etc.With the development of SAR imaging techniques,there has been a huge number of SAR images with high-resolution which were acquired for earth observation.Therefore,it is crucial to interpret these images quickly and accurately.SAR image scene classification is one of the key technologies in its interpretation.An excellent SAR image classification system is beneficial for fast and accurate SAR image interpretation.However,for SAR image scene classification,it is more crucial to extract discriminative features from SAR images than to learn an effective and robust classifier.To this end,in terms of the discriminative features extracted from SAR images,taking advantage of deep metric learning,the main research work is as follows.(1)As a result of heavy speckle,it is difficult to distinguish different SAR scene images effectively.In terms of this,a SAR image classification method based on the Mahalanobis distance is proposed by this thesis.Specifically,an embedding layer is added at the end of CNN,and the features extracted by CNN are mapped into the embedding space,and then a new objective function is introduced which consists of a cross-entropy loss and a triplet loss based on the Mahalanobis distance.During the training time,the parameters of CNN and the embedding layer can be learned at the same time by optimizing the objective function.In addition,an embedding space can be learned in which samples of the same classes are close to each other and samples of different classes are far away from each oth er.In the end,the discriminative feature extraction of SAR images is realized.And it will be applied to the SAR image scene classification in the test time.Numerous experiments on the SAR image scene classification indicate that this method outperforms other related methods in terms of Accuracy,Precision,Recall and F1-score.(2)With the increase of spatial resolution,the spatial structure of SAR images becomes very complicated,which makes it very difficult to describe.To solve this problem,with help of the previous work,this thesis proposes a SAR image classification method based on metric ensemble.Specifically,the embedding layer at the end of CNN is divided into multiple sub-embeddings,which share a common feature representation.A classifier is added at the end of each subembedding.At the training time,the sub-embedding and the classifier are learned by the online gradient boosting.Since SAR platforms are usually spaceborne or airborne,this enables SAR images to have inner-class diversity.To this end,a metric function based on the Pearson correlation coefficient is introduced to constrain sub-embeddings to increase the diversity of the embedding features.In the end,the complex scenes of SAR images can be described effectively,and the inner-class information of SAR images can be effectively utilized.The effectiveness of this method is proved by experiments.The experiment was verified by SAR dataset,and the method performance was evaluated by Accuracy,Precision,Recall and F1-score.Specifically,the experiment first compares other methods and then verifies this method by adjusting the number of sub-embeddings.The experiment proves that when the number of sub-embeddings is 4 and the size is 32,this method has the best effect in SAR image scene classification.In addition,the experiment also compares the proposed method with the SAR image classification method based on Mahalanobis distance.By comparison,it can be seen that the proposed method performs better than the latter in SAR image scene classification and effectively improves the classification accuracy.In conclusion,the SAR image scene classification method based on Mahalanobis distance proposed in this thesis effectively alleviates the influence of coherent spot and realizes the effective differentiation of different scenes.Experiments show that this method can effectively distinguish different scenes.Therefore,the SAR image classification method based on metric ensemble was designed to describe the complex spatial structure of SAR images,thus effectively utilizing the inner-class information of SAR images.Numerous experiments on the SAR dataset indicate that this method outperforms other related methods.