Research On Ocean Internal Waves Detection Based On Deep Learning In Remote Sensing Images

At present,the use of remote sensing satellites to observe ocean internal waves has become the main approach for researchers.Because manual interpretation of ocean internal waves is time-consuming and laborious,it cannot be used as an effective method for interpreting internal waves,and remote sensing data are rich and diverse.Therefore,it is necessary to develop an internal waves feature automatic detection technology to accelerate data processing through target feature extraction and target recognition technology.Deep learning has great advantages in target detection.This paper is based on deep learning to study ocean internal wave images detection.The main work of the thesis is:1.In this paper,the Faster R-CNN algorithm is used to study the ocean internal wave detection technology.The ZF-Net model is used to train the ocean internal waves data,and the internal wave morphological characteristics are learned to construct the internal wave detection network based on Faster R-CNN.2.In order to train the ocean wave images sample data and learn the internal wave shape characteristics through the convolutional network,this paper constructs a marine internal wave image database for the South China Sea region and the West Pacific region.For the South China Sea region,this paper uses the 2003-2012 ENVISAT ASAR data to construct the database of internal waves in the South China Sea,with a total of 946 samples.For the West Pacific area,this paper uses the 2017 Sentinel-1 data to construct the database of the Western Pacific region,with a total of 924 samples.3.We train the internal wave image sample data,adjust the network parameters and determine the accuracy threshold.When training the internal waves detection network,the detection results are optimized by training data ratio,learning rate and iteration numbers.Through a number of experiments,a set of data with the best detection effect is finally determined,and the setting of the accuracy threshold is determined by the FoM curve.By training the South China Sea region data,the final accuracy threshold is set to 0.33 and the FoM value is 0.90.By training the West Pacific region data,the final precision threshold is set to 0.20 and the FoM value is 0.90.4.This paper analyzes the internal waves of different shapes in different regions.For the South China Sea region and the Western Pacific region,the detection results of internal waves with different stripes and different scales are analyzed by the algorithm.For the internal solitary waves,wave packets and large-scale internal waves,the proposed algorithm not only has high detection accuracy,but also the detected region accurately surrounds the internal waves generating region.For wave packet groups and small-scale internal waves,the algorithm cannot accurately surround the internal wave region,but the detection accuracy is high.In addition,this paper examines the complex background features of ship wakes and fronts that are easily confused with ocean internal waves.The test results show that the algorithm does not detect the ship wakes and fronts as internal waves,which proves the effectiveness of the proposed algorithm.