Ship Classification In SAR Image Based On Remote Sensing Data Fusion

Synthetic aperture radar(SAR)is an active microwave sensor,can provide high-resolution,day-and-night and weather-independent images Since its birth in the 1950s,SAR has played an important role in ocean observation and ship monitoring.Ship classification is of great significance for maritime traffic management and safeguarding the country’s maritime rights and interests.It has been widely researched in recent years and has become an important study in the field of ship monitoring.The lack of labeled SAR ship data and how to extract more discriminative features are two general problems in the field of SAR ship image classification.Especially when deep learning is applied to SAR ship image classification,large-scale labeled samples are required to train the parameters in order to prevent over-fitting.At present,when deep learning is applied to small-scale datasets,the model usually be pre-trained on a large dataset.In SAR ship classification,considering the difference between the image in ImageNet and SAR ship image,the pre-trained network may not be suitable for extracting the feature of SAR ship Compared with SAR images,optical ship remote sensing images are not only easy to obtain,but also can be directly understood by human.Most importantly,optical ship images are similar to SAR ship images.In this paper,we hope to pre-train deep network in optical ship dataset to solve the problem of lack of labeled SAR ships,and extract more discriminative features.Therefore,we propose the method for ship classification in SAR image based on remote sensing data fusion.The main contributions in this paper can be summarized as follows1.The optical ship image dataset and the SAR ship image dataset are constructed,which contained 1093 high-resolution SAR ship data and 9072 optical ship data2.In order to solve the problem of lack of labeled SAR images,we pre-train the convolutional neural network(CNN)in optical ship dataset,and propose a deep feature method3.We compared 28 CNN network structures with different sizes and depths in three dataset.Through the analysis of experiment results and comparative experiments,the effectiveness of the proposed method is proved.4.We put forward expectations for future work,on the one hand,we will continue to expand the database,on the other hand,we will improve the proposed method in this paper and solve the problems in the experiment.