Automatic Recognition Of Mesoscale Eddy And Dynamic Relation With Fishing Ground Based On Ocean Remote Sensing Images

The ocean has become the main battlefield for resource competing in the world.And humans have entered the period of highly development and utilization of the ocean.Comprehensive understanding the ocean science is the key way to reasonable control,effective exploitation and sustainable development of the ocean.Mesoscale oceanic eddies are ubiquitous in global ocean and marginal seas as important ocean phenomena.Mesoscale oceanic eddies carry more than 90% of the ocean’s kinetic energy.They rotate and move horizontally in high speed with an irregular spiral structure by changing the vertical and horizontal distribution of the ocean energy and nutrients.Therefore,achieving the dynamics monitoring of the mesoscale oceanic eddies and analysis of characteristics of spatial and temporal variation is not only helpful to the study of marine ecological resource distribution,but also plays an important role in the practical application of deep sea fishing and ocean fishery.Automatic recognition of mesoscale oceanic eddies is the key way to achieve their dynamics monitoring and analysis of temporal and spatial variation characteristics.The ocean remote sensing provides an irreplaceable data source for automatic recognition research of mesoscale oceanic eddies.The ocean remote sensing can observe ocean phenomena and ocean environmental elements in remotely and non-contacting way and obtain rich remote sensing data.Mesoscale oceanic eddies recognition based on remote sensing data has become an advanced research hotspot.The main research methods can be divided into physical parameter-based,geometric-based and their hybrid methods.Physical parameters based methods need to set a suitable threshold for a specific region,they are therefore task-specific and limited in generalization capability.Geometrics-based methods are not able to detect eddies details from the data without clear geometrical features.However,the current methods which based on handcrafted feature introduces a large number of human subjective factors,resulting in the low accuracy of mesoscale eddies identification.In addition,compared with land remote sensing,ocean remote sensing images have significant weak features which are mainly manifested in the low spectral contrast and feature representation uncertainty of ocean phenomena caused by high dynamics.Meanwhile,mesoscale oceanic eddies have high natural variability caused by instability of even ocean current,coercive of sea surface wind,variety of bottom topography,and in the movement process with the energy injection or dissipation the geometric shape and physical properties are highly dynamic changes.The weak feature characteristic of ocean remote sensing and high dynamic of mesoscale oceanic eddies not only aggravate the limitations of exiting artificial feature design methods,but also the single threshold-based recognition method significantly lack generalization ability.The objective of this research is to realize the automatic identification of mesoscale oceanic eddies with high accuracy.Aiming at the limitation of the existing method for the automatic recognition of high dynamic mesoscale oceanic eddies and with the idea of deep learning,we put forward an automatic recognition model of mesoscale oceanic eddies based on feature learning.The main contents of this paper are as follows.(1)Build the mesoscale oceanic eddies training data set based on SAR images.SAR images obtained from radar sensors have a high spatial resolution,a wide swath of observation,and is not subject to cloud cover and sunlight conditions.Thus,SAR images are an effective source to gain more comprehensive and detailed information on mesoscale and sun-mesoscale phenomena in the oceans.The SAR images used in this paper are obtained by ESA-2 and Envisat in the 2005-2010 which provided by ESA.And the spatial range of the study sea area is 5°N-25°N,108°E-125°E.We annotated the mesoscale oceanic eddies by the bounding rectangle based on visual characteristics,and improve the volume of the training data set by data augmentation(2)Research on mesoscale oceanic eddies automatic recognition model based on feature learning.The acquisition and representation of the high level and essential features of mesoscale oceanic eddies is the key to realize their automatic recognition.Based on the small volume of the training data set in SAR image,we construct a multi-layer network model which suitable for high dynamic mesoscale oceanic eddies automatic recognition Deep Eddy from the model parameter initialization and model architecture these two aspects.Deep Eddy model can abstract the oceanic eddies high level feature layer-wise and recognize the oceanic eddies with high accuracy.Experiments show that the best recognition accuracy of the Deep Eddy model is 96.88%.(3)Propose a shape and scale robust automatic recognition model of mesoscale oceanic eddies.The mesoscale oceanic eddies have serious geometric deformation and space scale difference which significantly affect the accuracy of recognition.The acquisition of multi-scale spatial feature is the key to reduce the influence on the recognition accuracy which caused by geometric deformation and spatial scale difference.Based on the spatial pyramid model,the multi-layer network model of mesoscale oceanic eddies automatic recognition is improved,which is denoted as Deep Eddy+.The Deep Eddy+ realize the extraction and representation of mesoscale oceanic eddies multi-scale spatial feature.And the experiment results show that the recognition accuracy of Deep Eddy+ model is better than Deep Eddy model in the same model parameters setting.What’s more,the best recognition accuracy of Deep Eddy+ is 98.47%.(4)Empirical analysis of the automatic recognition model of mesoscale oceanic eddies.We recognize the mesoscale oceanic eddies in SAR images based on Deep Eddy+ model in study sea area.And the size,spatial and temporal variation characteristics of mesoscale oceanic eddies which recognized in SAR images are analyzed,and compared with the oceanic eddies which obtained based on SSH data recognition.The results show that the mesoscale oceanic eddies of the two data sources have significant differences in statistical characteristics and showing a complementary relationship.Additionally,this paper explores the spatial correlation of mesoscale oceanic eddies and tuna in the study sea area.The results show that there is a positive correlation between the bigeye tuna and cyclonic eddies,and the yellowfin tuna is highly correlated with the anti-cyclonic eddies.This paper has achieved some research results based on the above research.Specifically,(1)Build the mesoscale oceanic eddies training data set for the first time which provides a data basis for the future study of mesoscale oceanic eddies automatic recognition.(2)A simple and effective automatic recognition model of mesoscale oceanic eddies is proposed,which achieves the high recognition accuracy fully automated.(3)This paper provides a new theoretical method for the automatic recognition of high dynamic mesoscale oceanic eddies.At the same time,provide a technical reference for other marine phenomena recognition.