Detection Of Sea Motion Targets In Multi-spectral Imagery Of Static Orbiting Staring Satellites

Our country has a long coastline and vast sea area.The country pays more and more attention to the full utilization of marine resources.With the implementation of the strategy of the Maritime Silk Road,the country pays more attention to the full utilization of nearreal-time marine monitoring has become increasingly urgent.However,high-resolution synchronous trajectory remote sensing is still in its infancy in China,and the application of the only high-resolution remote sensing image in orbit has not yet been fully implemented.Therefore,based on the in-depth study of the imaging mechanism and characteristics of GF4,we systematically study the detection and motion estimation of sea moving targets based on GF4.The specific application of marine monitoring provides some reference.The details are as follows:First of all,in view of the problem that the foreground detection algorithm based on the visual attention mechanism cannot fully highlight the research object,we propose a series of feature descriptions that can highlight the sea surface motion target or suppress the main interference objects in the image,and use the local salience highlighting method to deal with some part of the features.And then we use the feature merging method to generate the comprehensive feature saliency map,thereby realizing the fast extraction of suspected sea surface motion target regions.Later,it is necessary to accurately identify the real target in the suspected area on the basis of the suspected target area,but the traditional methods have the disadvantages of large false alarm,high miss detection rate and complex calculation.Therefore,we propose a convolutional neural network target identification method based on migration learning to accurate identify the real target,by training the network with self-defined feature channels.Furthermore,by using the identification model to identify the suspected area,we concatenate the unsupervised fast extraction algorithm of the suspected area and the supervised target accurate identification algorithm,which constitute the cascade sea moving targets detection algorithm.It can realize the fast positioning of the moving targets in the image of GF4.Finally,according to the time-sharing imaging mechanism of GF4,we design and implement the motion state estimation method of detected target based on image,by combining the displacement of moving target in different channels of single scene image,with the imaging interval between band and band,which successfully estimated the speed and course of the target.By establishing the registration relation between image and ground measured data source,we compare the experimental results with the ground measured data,and analyze the reliability and limitation of the estimation method.