Deep Learning-based Ocean Internal Waves Detection From SAR Images In South China Sea

Ocean internal waves are widely present in oceans and marginal sea areas.transferring large and medium-scale energy to small-scale energy by nonlinear motion process.Ocean internal waves play an important role in seawater energy exchange and are safety risks for offshore production,submarine navigation,etc.As the largest marginal sea of China,the South China Sea is crucial in geopolitics and rich in resources.However,it is also one of the sea areas where internal ocean waves occur most frequently.The study of ocean internal waves has important academic value and practical significance in marine resources,marine engineering,and the marine military.Ocean satellite remote sensing technology can quickly obtain information about ocean phenomena and ocean environment elements from a long distance,without contact,and on a large scale,which provides a database for ocean research.Synthetic aperture radar(SAR)is an indispensable data source for Ocean internal waves research,as can provide full-time and fullweather observation for targets.with the accumulation of data under the long-term series,the SAR data shows the "5V" big data characteristics.The existing automatic detection methods of ocean internal waves are mainly based on geometric features for identification.The subjectivity of manually designed features resulted in low recognition accuracy.Besides,other features presented in SAR images,such as ship tail,sea surface oil spill,and oceanfront surface,are easily detected as ocean internal waves,which leads to a low recognition accuracy of the existed detection algorithms.Deep learning brings new opportunities for the automatic detection of ocean internal waves in SAR images.The ocean internal waves are irregular light and dark stripes in SAR images.Based on these features,a deep learning-based method is developed for ocean internal waves detection in multi-band SAR images.The main contents and results of the thesis are as follows.5480 SAR images are collected as raining data set from 2001 to 2020 in the South China Sea region,which is in multi-band(C?L?X),multi-polarization mode,and multi-spatial scale.The developed ocean internal waves are based on the Fast R-CNN network framework,combined with the transfer learning method in the training process.The accuracy rate(AP)and recall rate(AR)is up to 0.957 and 0.923 respectively,and the average detection speed is 5fps.the methods are evaluated by multi-source SAR data and the ocean internal waves are in different scenarios.The results show that:(1)the detection algorithm has good performance on SAR images with multi-band,different polarization mode and different spatial resolution;(2)the algorithm shows low misdetection rate for ocean internal waves in for different spatial scale and complex forms;and(3)for SAR images with a complex background,this algorithm also shows resistant to interference.Base on the above-mentioned research,the main contribution of this thesis are:(1)developed a training data set of ocean internal waves based on SAR images,which provides a data basis for the research of automatic detection of ocean internal waves;(2)designed an effective model for automatic detecting ocean internal waves based on deep learning;and(3)offered a potential avenue for the automatic identification of other ocean dynamic process and phenomena.